2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
66 #include <asm/sections.h>
67 #include <asm/tlbflush.h>
68 #include <asm/div64.h>
71 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
72 static DEFINE_MUTEX(pcp_batch_high_lock);
73 #define MIN_PERCPU_PAGELIST_FRACTION (8)
75 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
76 DEFINE_PER_CPU(int, numa_node);
77 EXPORT_PER_CPU_SYMBOL(numa_node);
80 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
82 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
83 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
84 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
85 * defined in <linux/topology.h>.
87 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
88 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
89 int _node_numa_mem_[MAX_NUMNODES];
93 * Array of node states.
95 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
96 [N_POSSIBLE] = NODE_MASK_ALL,
97 [N_ONLINE] = { { [0] = 1UL } },
99 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 #ifdef CONFIG_HIGHMEM
101 [N_HIGH_MEMORY] = { { [0] = 1UL } },
103 #ifdef CONFIG_MOVABLE_NODE
104 [N_MEMORY] = { { [0] = 1UL } },
106 [N_CPU] = { { [0] = 1UL } },
109 EXPORT_SYMBOL(node_states);
111 /* Protect totalram_pages and zone->managed_pages */
112 static DEFINE_SPINLOCK(managed_page_count_lock);
114 unsigned long totalram_pages __read_mostly;
115 unsigned long totalreserve_pages __read_mostly;
116 unsigned long totalcma_pages __read_mostly;
118 int percpu_pagelist_fraction;
119 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
122 * A cached value of the page's pageblock's migratetype, used when the page is
123 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
124 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
125 * Also the migratetype set in the page does not necessarily match the pcplist
126 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
127 * other index - this ensures that it will be put on the correct CMA freelist.
129 static inline int get_pcppage_migratetype(struct page *page)
134 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
136 page->index = migratetype;
139 #ifdef CONFIG_PM_SLEEP
141 * The following functions are used by the suspend/hibernate code to temporarily
142 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
143 * while devices are suspended. To avoid races with the suspend/hibernate code,
144 * they should always be called with pm_mutex held (gfp_allowed_mask also should
145 * only be modified with pm_mutex held, unless the suspend/hibernate code is
146 * guaranteed not to run in parallel with that modification).
149 static gfp_t saved_gfp_mask;
151 void pm_restore_gfp_mask(void)
153 WARN_ON(!mutex_is_locked(&pm_mutex));
154 if (saved_gfp_mask) {
155 gfp_allowed_mask = saved_gfp_mask;
160 void pm_restrict_gfp_mask(void)
162 WARN_ON(!mutex_is_locked(&pm_mutex));
163 WARN_ON(saved_gfp_mask);
164 saved_gfp_mask = gfp_allowed_mask;
165 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
168 bool pm_suspended_storage(void)
170 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
174 #endif /* CONFIG_PM_SLEEP */
176 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
177 unsigned int pageblock_order __read_mostly;
180 static void __free_pages_ok(struct page *page, unsigned int order);
183 * results with 256, 32 in the lowmem_reserve sysctl:
184 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
185 * 1G machine -> (16M dma, 784M normal, 224M high)
186 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
187 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
188 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
190 * TBD: should special case ZONE_DMA32 machines here - in those we normally
191 * don't need any ZONE_NORMAL reservation
193 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
200 #ifdef CONFIG_HIGHMEM
206 EXPORT_SYMBOL(totalram_pages);
208 static char * const zone_names[MAX_NR_ZONES] = {
209 #ifdef CONFIG_ZONE_DMA
212 #ifdef CONFIG_ZONE_DMA32
216 #ifdef CONFIG_HIGHMEM
220 #ifdef CONFIG_ZONE_DEVICE
225 static void free_compound_page(struct page *page);
226 compound_page_dtor * const compound_page_dtors[] = {
229 #ifdef CONFIG_HUGETLB_PAGE
235 * Try to keep at least this much lowmem free. Do not allow normal
236 * allocations below this point, only high priority ones. Automatically
237 * tuned according to the amount of memory in the system.
239 int min_free_kbytes = 1024;
240 int user_min_free_kbytes = -1;
243 * Extra memory for the system to try freeing. Used to temporarily
244 * free memory, to make space for new workloads. Anyone can allocate
245 * down to the min watermarks controlled by min_free_kbytes above.
247 int extra_free_kbytes = 0;
249 static unsigned long __meminitdata nr_kernel_pages;
250 static unsigned long __meminitdata nr_all_pages;
251 static unsigned long __meminitdata dma_reserve;
253 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
254 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
255 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
256 static unsigned long __initdata required_kernelcore;
257 static unsigned long __initdata required_movablecore;
258 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
260 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
262 EXPORT_SYMBOL(movable_zone);
263 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
266 int nr_node_ids __read_mostly = MAX_NUMNODES;
267 int nr_online_nodes __read_mostly = 1;
268 EXPORT_SYMBOL(nr_node_ids);
269 EXPORT_SYMBOL(nr_online_nodes);
272 int page_group_by_mobility_disabled __read_mostly;
274 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
275 static inline void reset_deferred_meminit(pg_data_t *pgdat)
277 unsigned long max_initialise;
278 unsigned long reserved_lowmem;
281 * Initialise at least 2G of a node but also take into account that
282 * two large system hashes that can take up 1GB for 0.25TB/node.
284 max_initialise = max(2UL << (30 - PAGE_SHIFT),
285 (pgdat->node_spanned_pages >> 8));
288 * Compensate the all the memblock reservations (e.g. crash kernel)
289 * from the initial estimation to make sure we will initialize enough
292 reserved_lowmem = memblock_reserved_memory_within(pgdat->node_start_pfn,
293 pgdat->node_start_pfn + max_initialise);
294 max_initialise += reserved_lowmem;
296 pgdat->static_init_size = min(max_initialise, pgdat->node_spanned_pages);
297 pgdat->first_deferred_pfn = ULONG_MAX;
300 /* Returns true if the struct page for the pfn is uninitialised */
301 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
303 int nid = early_pfn_to_nid(pfn);
305 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
311 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
313 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
320 * Returns false when the remaining initialisation should be deferred until
321 * later in the boot cycle when it can be parallelised.
323 static inline bool update_defer_init(pg_data_t *pgdat,
324 unsigned long pfn, unsigned long zone_end,
325 unsigned long *nr_initialised)
327 /* Always populate low zones for address-contrained allocations */
328 if (zone_end < pgdat_end_pfn(pgdat))
330 /* Initialise at least 2G of the highest zone */
332 if ((*nr_initialised > pgdat->static_init_size) &&
333 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
334 pgdat->first_deferred_pfn = pfn;
341 static inline void reset_deferred_meminit(pg_data_t *pgdat)
345 static inline bool early_page_uninitialised(unsigned long pfn)
350 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
355 static inline bool update_defer_init(pg_data_t *pgdat,
356 unsigned long pfn, unsigned long zone_end,
357 unsigned long *nr_initialised)
364 void set_pageblock_migratetype(struct page *page, int migratetype)
366 if (unlikely(page_group_by_mobility_disabled &&
367 migratetype < MIGRATE_PCPTYPES))
368 migratetype = MIGRATE_UNMOVABLE;
370 set_pageblock_flags_group(page, (unsigned long)migratetype,
371 PB_migrate, PB_migrate_end);
374 #ifdef CONFIG_DEBUG_VM
375 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
379 unsigned long pfn = page_to_pfn(page);
380 unsigned long sp, start_pfn;
383 seq = zone_span_seqbegin(zone);
384 start_pfn = zone->zone_start_pfn;
385 sp = zone->spanned_pages;
386 if (!zone_spans_pfn(zone, pfn))
388 } while (zone_span_seqretry(zone, seq));
391 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
392 pfn, zone_to_nid(zone), zone->name,
393 start_pfn, start_pfn + sp);
398 static int page_is_consistent(struct zone *zone, struct page *page)
400 if (!pfn_valid_within(page_to_pfn(page)))
402 if (zone != page_zone(page))
408 * Temporary debugging check for pages not lying within a given zone.
410 static int bad_range(struct zone *zone, struct page *page)
412 if (page_outside_zone_boundaries(zone, page))
414 if (!page_is_consistent(zone, page))
420 static inline int bad_range(struct zone *zone, struct page *page)
426 static void bad_page(struct page *page, const char *reason,
427 unsigned long bad_flags)
429 static unsigned long resume;
430 static unsigned long nr_shown;
431 static unsigned long nr_unshown;
433 /* Don't complain about poisoned pages */
434 if (PageHWPoison(page)) {
435 page_mapcount_reset(page); /* remove PageBuddy */
440 * Allow a burst of 60 reports, then keep quiet for that minute;
441 * or allow a steady drip of one report per second.
443 if (nr_shown == 60) {
444 if (time_before(jiffies, resume)) {
450 "BUG: Bad page state: %lu messages suppressed\n",
457 resume = jiffies + 60 * HZ;
459 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
460 current->comm, page_to_pfn(page));
461 dump_page_badflags(page, reason, bad_flags);
466 /* Leave bad fields for debug, except PageBuddy could make trouble */
467 page_mapcount_reset(page); /* remove PageBuddy */
468 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
472 * Higher-order pages are called "compound pages". They are structured thusly:
474 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
476 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
477 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
479 * The first tail page's ->compound_dtor holds the offset in array of compound
480 * page destructors. See compound_page_dtors.
482 * The first tail page's ->compound_order holds the order of allocation.
483 * This usage means that zero-order pages may not be compound.
486 static void free_compound_page(struct page *page)
488 __free_pages_ok(page, compound_order(page));
491 void prep_compound_page(struct page *page, unsigned int order)
494 int nr_pages = 1 << order;
496 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
497 set_compound_order(page, order);
499 for (i = 1; i < nr_pages; i++) {
500 struct page *p = page + i;
501 set_page_count(p, 0);
502 set_compound_head(p, page);
506 #ifdef CONFIG_DEBUG_PAGEALLOC
507 unsigned int _debug_guardpage_minorder;
508 bool _debug_pagealloc_enabled __read_mostly;
509 bool _debug_guardpage_enabled __read_mostly;
511 static int __init early_debug_pagealloc(char *buf)
516 if (strcmp(buf, "on") == 0)
517 _debug_pagealloc_enabled = true;
521 early_param("debug_pagealloc", early_debug_pagealloc);
523 static bool need_debug_guardpage(void)
525 /* If we don't use debug_pagealloc, we don't need guard page */
526 if (!debug_pagealloc_enabled())
532 static void init_debug_guardpage(void)
534 if (!debug_pagealloc_enabled())
537 _debug_guardpage_enabled = true;
540 struct page_ext_operations debug_guardpage_ops = {
541 .need = need_debug_guardpage,
542 .init = init_debug_guardpage,
545 static int __init debug_guardpage_minorder_setup(char *buf)
549 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
550 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
553 _debug_guardpage_minorder = res;
554 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
557 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
559 static inline void set_page_guard(struct zone *zone, struct page *page,
560 unsigned int order, int migratetype)
562 struct page_ext *page_ext;
564 if (!debug_guardpage_enabled())
567 page_ext = lookup_page_ext(page);
568 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
570 INIT_LIST_HEAD(&page->lru);
571 set_page_private(page, order);
572 /* Guard pages are not available for any usage */
573 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
576 static inline void clear_page_guard(struct zone *zone, struct page *page,
577 unsigned int order, int migratetype)
579 struct page_ext *page_ext;
581 if (!debug_guardpage_enabled())
584 page_ext = lookup_page_ext(page);
585 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
587 set_page_private(page, 0);
588 if (!is_migrate_isolate(migratetype))
589 __mod_zone_freepage_state(zone, (1 << order), migratetype);
592 struct page_ext_operations debug_guardpage_ops = { NULL, };
593 static inline void set_page_guard(struct zone *zone, struct page *page,
594 unsigned int order, int migratetype) {}
595 static inline void clear_page_guard(struct zone *zone, struct page *page,
596 unsigned int order, int migratetype) {}
599 static inline void set_page_order(struct page *page, unsigned int order)
601 set_page_private(page, order);
602 __SetPageBuddy(page);
605 static inline void rmv_page_order(struct page *page)
607 __ClearPageBuddy(page);
608 set_page_private(page, 0);
612 * This function checks whether a page is free && is the buddy
613 * we can do coalesce a page and its buddy if
614 * (a) the buddy is not in a hole &&
615 * (b) the buddy is in the buddy system &&
616 * (c) a page and its buddy have the same order &&
617 * (d) a page and its buddy are in the same zone.
619 * For recording whether a page is in the buddy system, we set ->_mapcount
620 * PAGE_BUDDY_MAPCOUNT_VALUE.
621 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
622 * serialized by zone->lock.
624 * For recording page's order, we use page_private(page).
626 static inline int page_is_buddy(struct page *page, struct page *buddy,
629 if (!pfn_valid_within(page_to_pfn(buddy)))
632 if (page_is_guard(buddy) && page_order(buddy) == order) {
633 if (page_zone_id(page) != page_zone_id(buddy))
636 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
641 if (PageBuddy(buddy) && page_order(buddy) == order) {
643 * zone check is done late to avoid uselessly
644 * calculating zone/node ids for pages that could
647 if (page_zone_id(page) != page_zone_id(buddy))
650 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
658 * Freeing function for a buddy system allocator.
660 * The concept of a buddy system is to maintain direct-mapped table
661 * (containing bit values) for memory blocks of various "orders".
662 * The bottom level table contains the map for the smallest allocatable
663 * units of memory (here, pages), and each level above it describes
664 * pairs of units from the levels below, hence, "buddies".
665 * At a high level, all that happens here is marking the table entry
666 * at the bottom level available, and propagating the changes upward
667 * as necessary, plus some accounting needed to play nicely with other
668 * parts of the VM system.
669 * At each level, we keep a list of pages, which are heads of continuous
670 * free pages of length of (1 << order) and marked with _mapcount
671 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
673 * So when we are allocating or freeing one, we can derive the state of the
674 * other. That is, if we allocate a small block, and both were
675 * free, the remainder of the region must be split into blocks.
676 * If a block is freed, and its buddy is also free, then this
677 * triggers coalescing into a block of larger size.
682 static inline void __free_one_page(struct page *page,
684 struct zone *zone, unsigned int order,
687 unsigned long page_idx;
688 unsigned long combined_idx;
689 unsigned long uninitialized_var(buddy_idx);
691 unsigned int max_order;
693 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
695 VM_BUG_ON(!zone_is_initialized(zone));
696 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
698 VM_BUG_ON(migratetype == -1);
699 if (likely(!is_migrate_isolate(migratetype)))
700 __mod_zone_freepage_state(zone, 1 << order, migratetype);
702 page_idx = pfn & ((1 << MAX_ORDER) - 1);
704 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
705 VM_BUG_ON_PAGE(bad_range(zone, page), page);
708 while (order < max_order - 1) {
709 buddy_idx = __find_buddy_index(page_idx, order);
710 buddy = page + (buddy_idx - page_idx);
711 if (!page_is_buddy(page, buddy, order))
714 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
715 * merge with it and move up one order.
717 if (page_is_guard(buddy)) {
718 clear_page_guard(zone, buddy, order, migratetype);
720 list_del(&buddy->lru);
721 zone->free_area[order].nr_free--;
722 rmv_page_order(buddy);
724 combined_idx = buddy_idx & page_idx;
725 page = page + (combined_idx - page_idx);
726 page_idx = combined_idx;
729 if (max_order < MAX_ORDER) {
730 /* If we are here, it means order is >= pageblock_order.
731 * We want to prevent merge between freepages on isolate
732 * pageblock and normal pageblock. Without this, pageblock
733 * isolation could cause incorrect freepage or CMA accounting.
735 * We don't want to hit this code for the more frequent
738 if (unlikely(has_isolate_pageblock(zone))) {
741 buddy_idx = __find_buddy_index(page_idx, order);
742 buddy = page + (buddy_idx - page_idx);
743 buddy_mt = get_pageblock_migratetype(buddy);
745 if (migratetype != buddy_mt
746 && (is_migrate_isolate(migratetype) ||
747 is_migrate_isolate(buddy_mt)))
751 goto continue_merging;
755 set_page_order(page, order);
758 * If this is not the largest possible page, check if the buddy
759 * of the next-highest order is free. If it is, it's possible
760 * that pages are being freed that will coalesce soon. In case,
761 * that is happening, add the free page to the tail of the list
762 * so it's less likely to be used soon and more likely to be merged
763 * as a higher order page
765 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
766 struct page *higher_page, *higher_buddy;
767 combined_idx = buddy_idx & page_idx;
768 higher_page = page + (combined_idx - page_idx);
769 buddy_idx = __find_buddy_index(combined_idx, order + 1);
770 higher_buddy = higher_page + (buddy_idx - combined_idx);
771 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
772 list_add_tail(&page->lru,
773 &zone->free_area[order].free_list[migratetype]);
778 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
780 zone->free_area[order].nr_free++;
783 static inline int free_pages_check(struct page *page)
785 const char *bad_reason = NULL;
786 unsigned long bad_flags = 0;
788 if (unlikely(page_mapcount(page)))
789 bad_reason = "nonzero mapcount";
790 if (unlikely(page->mapping != NULL))
791 bad_reason = "non-NULL mapping";
792 if (unlikely(atomic_read(&page->_count) != 0))
793 bad_reason = "nonzero _count";
794 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
795 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
796 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
799 if (unlikely(page->mem_cgroup))
800 bad_reason = "page still charged to cgroup";
802 if (unlikely(bad_reason)) {
803 bad_page(page, bad_reason, bad_flags);
806 page_cpupid_reset_last(page);
807 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
808 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
813 * Frees a number of pages from the PCP lists
814 * Assumes all pages on list are in same zone, and of same order.
815 * count is the number of pages to free.
817 * If the zone was previously in an "all pages pinned" state then look to
818 * see if this freeing clears that state.
820 * And clear the zone's pages_scanned counter, to hold off the "all pages are
821 * pinned" detection logic.
823 static void free_pcppages_bulk(struct zone *zone, int count,
824 struct per_cpu_pages *pcp)
829 unsigned long nr_scanned;
831 spin_lock(&zone->lock);
832 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
834 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
838 struct list_head *list;
841 * Remove pages from lists in a round-robin fashion. A
842 * batch_free count is maintained that is incremented when an
843 * empty list is encountered. This is so more pages are freed
844 * off fuller lists instead of spinning excessively around empty
849 if (++migratetype == MIGRATE_PCPTYPES)
851 list = &pcp->lists[migratetype];
852 } while (list_empty(list));
854 /* This is the only non-empty list. Free them all. */
855 if (batch_free == MIGRATE_PCPTYPES)
856 batch_free = to_free;
859 int mt; /* migratetype of the to-be-freed page */
861 page = list_entry(list->prev, struct page, lru);
862 /* must delete as __free_one_page list manipulates */
863 list_del(&page->lru);
865 mt = get_pcppage_migratetype(page);
866 /* MIGRATE_ISOLATE page should not go to pcplists */
867 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
868 /* Pageblock could have been isolated meanwhile */
869 if (unlikely(has_isolate_pageblock(zone)))
870 mt = get_pageblock_migratetype(page);
872 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
873 trace_mm_page_pcpu_drain(page, 0, mt);
874 } while (--to_free && --batch_free && !list_empty(list));
876 spin_unlock(&zone->lock);
879 static void free_one_page(struct zone *zone,
880 struct page *page, unsigned long pfn,
884 unsigned long nr_scanned;
885 spin_lock(&zone->lock);
886 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
888 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
890 if (unlikely(has_isolate_pageblock(zone) ||
891 is_migrate_isolate(migratetype))) {
892 migratetype = get_pfnblock_migratetype(page, pfn);
894 __free_one_page(page, pfn, zone, order, migratetype);
895 spin_unlock(&zone->lock);
898 static int free_tail_pages_check(struct page *head_page, struct page *page)
903 * We rely page->lru.next never has bit 0 set, unless the page
904 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
906 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
908 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
912 if (unlikely(!PageTail(page))) {
913 bad_page(page, "PageTail not set", 0);
916 if (unlikely(compound_head(page) != head_page)) {
917 bad_page(page, "compound_head not consistent", 0);
922 clear_compound_head(page);
926 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
927 unsigned long zone, int nid)
929 set_page_links(page, zone, nid, pfn);
930 init_page_count(page);
931 page_mapcount_reset(page);
932 page_cpupid_reset_last(page);
934 INIT_LIST_HEAD(&page->lru);
935 #ifdef WANT_PAGE_VIRTUAL
936 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
937 if (!is_highmem_idx(zone))
938 set_page_address(page, __va(pfn << PAGE_SHIFT));
942 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
945 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
948 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
949 static void init_reserved_page(unsigned long pfn)
954 if (!early_page_uninitialised(pfn))
957 nid = early_pfn_to_nid(pfn);
958 pgdat = NODE_DATA(nid);
960 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
961 struct zone *zone = &pgdat->node_zones[zid];
963 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
966 __init_single_pfn(pfn, zid, nid);
969 static inline void init_reserved_page(unsigned long pfn)
972 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
975 * Initialised pages do not have PageReserved set. This function is
976 * called for each range allocated by the bootmem allocator and
977 * marks the pages PageReserved. The remaining valid pages are later
978 * sent to the buddy page allocator.
980 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
982 unsigned long start_pfn = PFN_DOWN(start);
983 unsigned long end_pfn = PFN_UP(end);
985 for (; start_pfn < end_pfn; start_pfn++) {
986 if (pfn_valid(start_pfn)) {
987 struct page *page = pfn_to_page(start_pfn);
989 init_reserved_page(start_pfn);
991 /* Avoid false-positive PageTail() */
992 INIT_LIST_HEAD(&page->lru);
994 SetPageReserved(page);
999 static bool free_pages_prepare(struct page *page, unsigned int order)
1001 bool compound = PageCompound(page);
1004 VM_BUG_ON_PAGE(PageTail(page), page);
1005 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1007 trace_mm_page_free(page, order);
1008 kmemcheck_free_shadow(page, order);
1009 kasan_free_pages(page, order);
1012 page->mapping = NULL;
1013 bad += free_pages_check(page);
1014 for (i = 1; i < (1 << order); i++) {
1016 bad += free_tail_pages_check(page, page + i);
1017 bad += free_pages_check(page + i);
1022 reset_page_owner(page, order);
1024 if (!PageHighMem(page)) {
1025 debug_check_no_locks_freed(page_address(page),
1026 PAGE_SIZE << order);
1027 debug_check_no_obj_freed(page_address(page),
1028 PAGE_SIZE << order);
1030 arch_free_page(page, order);
1031 kernel_map_pages(page, 1 << order, 0);
1036 static void __free_pages_ok(struct page *page, unsigned int order)
1038 unsigned long flags;
1040 unsigned long pfn = page_to_pfn(page);
1042 if (!free_pages_prepare(page, order))
1045 migratetype = get_pfnblock_migratetype(page, pfn);
1046 local_irq_save(flags);
1047 __count_vm_events(PGFREE, 1 << order);
1048 free_one_page(page_zone(page), page, pfn, order, migratetype);
1049 local_irq_restore(flags);
1052 static void __init __free_pages_boot_core(struct page *page,
1053 unsigned long pfn, unsigned int order)
1055 unsigned int nr_pages = 1 << order;
1056 struct page *p = page;
1060 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1062 __ClearPageReserved(p);
1063 set_page_count(p, 0);
1065 __ClearPageReserved(p);
1066 set_page_count(p, 0);
1068 page_zone(page)->managed_pages += nr_pages;
1069 set_page_refcounted(page);
1070 __free_pages(page, order);
1073 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1074 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1076 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1078 int __meminit early_pfn_to_nid(unsigned long pfn)
1080 static DEFINE_SPINLOCK(early_pfn_lock);
1083 spin_lock(&early_pfn_lock);
1084 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1086 nid = first_online_node;
1087 spin_unlock(&early_pfn_lock);
1093 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1094 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1095 struct mminit_pfnnid_cache *state)
1099 nid = __early_pfn_to_nid(pfn, state);
1100 if (nid >= 0 && nid != node)
1105 /* Only safe to use early in boot when initialisation is single-threaded */
1106 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1108 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1113 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1117 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1118 struct mminit_pfnnid_cache *state)
1125 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1128 if (early_page_uninitialised(pfn))
1130 return __free_pages_boot_core(page, pfn, order);
1133 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1134 static void __init deferred_free_range(struct page *page,
1135 unsigned long pfn, int nr_pages)
1142 /* Free a large naturally-aligned chunk if possible */
1143 if (nr_pages == MAX_ORDER_NR_PAGES &&
1144 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1145 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1146 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1150 for (i = 0; i < nr_pages; i++, page++, pfn++)
1151 __free_pages_boot_core(page, pfn, 0);
1154 /* Completion tracking for deferred_init_memmap() threads */
1155 static atomic_t pgdat_init_n_undone __initdata;
1156 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1158 static inline void __init pgdat_init_report_one_done(void)
1160 if (atomic_dec_and_test(&pgdat_init_n_undone))
1161 complete(&pgdat_init_all_done_comp);
1164 /* Initialise remaining memory on a node */
1165 static int __init deferred_init_memmap(void *data)
1167 pg_data_t *pgdat = data;
1168 int nid = pgdat->node_id;
1169 struct mminit_pfnnid_cache nid_init_state = { };
1170 unsigned long start = jiffies;
1171 unsigned long nr_pages = 0;
1172 unsigned long walk_start, walk_end;
1175 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1176 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1178 if (first_init_pfn == ULONG_MAX) {
1179 pgdat_init_report_one_done();
1183 /* Bind memory initialisation thread to a local node if possible */
1184 if (!cpumask_empty(cpumask))
1185 set_cpus_allowed_ptr(current, cpumask);
1187 /* Sanity check boundaries */
1188 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1189 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1190 pgdat->first_deferred_pfn = ULONG_MAX;
1192 /* Only the highest zone is deferred so find it */
1193 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1194 zone = pgdat->node_zones + zid;
1195 if (first_init_pfn < zone_end_pfn(zone))
1199 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1200 unsigned long pfn, end_pfn;
1201 struct page *page = NULL;
1202 struct page *free_base_page = NULL;
1203 unsigned long free_base_pfn = 0;
1206 end_pfn = min(walk_end, zone_end_pfn(zone));
1207 pfn = first_init_pfn;
1208 if (pfn < walk_start)
1210 if (pfn < zone->zone_start_pfn)
1211 pfn = zone->zone_start_pfn;
1213 for (; pfn < end_pfn; pfn++) {
1214 if (!pfn_valid_within(pfn))
1218 * Ensure pfn_valid is checked every
1219 * MAX_ORDER_NR_PAGES for memory holes
1221 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1222 if (!pfn_valid(pfn)) {
1228 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1233 /* Minimise pfn page lookups and scheduler checks */
1234 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1237 nr_pages += nr_to_free;
1238 deferred_free_range(free_base_page,
1239 free_base_pfn, nr_to_free);
1240 free_base_page = NULL;
1241 free_base_pfn = nr_to_free = 0;
1243 page = pfn_to_page(pfn);
1248 VM_BUG_ON(page_zone(page) != zone);
1252 __init_single_page(page, pfn, zid, nid);
1253 if (!free_base_page) {
1254 free_base_page = page;
1255 free_base_pfn = pfn;
1260 /* Where possible, batch up pages for a single free */
1263 /* Free the current block of pages to allocator */
1264 nr_pages += nr_to_free;
1265 deferred_free_range(free_base_page, free_base_pfn,
1267 free_base_page = NULL;
1268 free_base_pfn = nr_to_free = 0;
1271 first_init_pfn = max(end_pfn, first_init_pfn);
1274 /* Sanity check that the next zone really is unpopulated */
1275 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1277 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1278 jiffies_to_msecs(jiffies - start));
1280 pgdat_init_report_one_done();
1284 void __init page_alloc_init_late(void)
1288 /* There will be num_node_state(N_MEMORY) threads */
1289 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1290 for_each_node_state(nid, N_MEMORY) {
1291 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1294 /* Block until all are initialised */
1295 wait_for_completion(&pgdat_init_all_done_comp);
1297 /* Reinit limits that are based on free pages after the kernel is up */
1298 files_maxfiles_init();
1300 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1303 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1304 void __init init_cma_reserved_pageblock(struct page *page)
1306 unsigned i = pageblock_nr_pages;
1307 struct page *p = page;
1310 __ClearPageReserved(p);
1311 set_page_count(p, 0);
1314 set_pageblock_migratetype(page, MIGRATE_CMA);
1316 if (pageblock_order >= MAX_ORDER) {
1317 i = pageblock_nr_pages;
1320 set_page_refcounted(p);
1321 __free_pages(p, MAX_ORDER - 1);
1322 p += MAX_ORDER_NR_PAGES;
1323 } while (i -= MAX_ORDER_NR_PAGES);
1325 set_page_refcounted(page);
1326 __free_pages(page, pageblock_order);
1329 adjust_managed_page_count(page, pageblock_nr_pages);
1334 * The order of subdivision here is critical for the IO subsystem.
1335 * Please do not alter this order without good reasons and regression
1336 * testing. Specifically, as large blocks of memory are subdivided,
1337 * the order in which smaller blocks are delivered depends on the order
1338 * they're subdivided in this function. This is the primary factor
1339 * influencing the order in which pages are delivered to the IO
1340 * subsystem according to empirical testing, and this is also justified
1341 * by considering the behavior of a buddy system containing a single
1342 * large block of memory acted on by a series of small allocations.
1343 * This behavior is a critical factor in sglist merging's success.
1347 static inline void expand(struct zone *zone, struct page *page,
1348 int low, int high, struct free_area *area,
1351 unsigned long size = 1 << high;
1353 while (high > low) {
1357 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1359 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1360 debug_guardpage_enabled() &&
1361 high < debug_guardpage_minorder()) {
1363 * Mark as guard pages (or page), that will allow to
1364 * merge back to allocator when buddy will be freed.
1365 * Corresponding page table entries will not be touched,
1366 * pages will stay not present in virtual address space
1368 set_page_guard(zone, &page[size], high, migratetype);
1371 list_add(&page[size].lru, &area->free_list[migratetype]);
1373 set_page_order(&page[size], high);
1378 * This page is about to be returned from the page allocator
1380 static inline int check_new_page(struct page *page)
1382 const char *bad_reason = NULL;
1383 unsigned long bad_flags = 0;
1385 if (unlikely(page_mapcount(page)))
1386 bad_reason = "nonzero mapcount";
1387 if (unlikely(page->mapping != NULL))
1388 bad_reason = "non-NULL mapping";
1389 if (unlikely(atomic_read(&page->_count) != 0))
1390 bad_reason = "nonzero _count";
1391 if (unlikely(page->flags & __PG_HWPOISON)) {
1392 bad_reason = "HWPoisoned (hardware-corrupted)";
1393 bad_flags = __PG_HWPOISON;
1395 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1396 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1397 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1400 if (unlikely(page->mem_cgroup))
1401 bad_reason = "page still charged to cgroup";
1403 if (unlikely(bad_reason)) {
1404 bad_page(page, bad_reason, bad_flags);
1410 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1415 for (i = 0; i < (1 << order); i++) {
1416 struct page *p = page + i;
1417 if (unlikely(check_new_page(p)))
1421 set_page_private(page, 0);
1422 set_page_refcounted(page);
1424 arch_alloc_page(page, order);
1425 kernel_map_pages(page, 1 << order, 1);
1426 kasan_alloc_pages(page, order);
1428 if (gfp_flags & __GFP_ZERO)
1429 for (i = 0; i < (1 << order); i++)
1430 clear_highpage(page + i);
1432 if (order && (gfp_flags & __GFP_COMP))
1433 prep_compound_page(page, order);
1435 set_page_owner(page, order, gfp_flags);
1438 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1439 * allocate the page. The expectation is that the caller is taking
1440 * steps that will free more memory. The caller should avoid the page
1441 * being used for !PFMEMALLOC purposes.
1443 if (alloc_flags & ALLOC_NO_WATERMARKS)
1444 set_page_pfmemalloc(page);
1446 clear_page_pfmemalloc(page);
1452 * Go through the free lists for the given migratetype and remove
1453 * the smallest available page from the freelists
1456 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1459 unsigned int current_order;
1460 struct free_area *area;
1463 /* Find a page of the appropriate size in the preferred list */
1464 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1465 area = &(zone->free_area[current_order]);
1466 if (list_empty(&area->free_list[migratetype]))
1469 page = list_entry(area->free_list[migratetype].next,
1471 list_del(&page->lru);
1472 rmv_page_order(page);
1474 expand(zone, page, order, current_order, area, migratetype);
1475 set_pcppage_migratetype(page, migratetype);
1484 * This array describes the order lists are fallen back to when
1485 * the free lists for the desirable migrate type are depleted
1487 static int fallbacks[MIGRATE_TYPES][4] = {
1488 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1489 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1490 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1492 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1494 #ifdef CONFIG_MEMORY_ISOLATION
1495 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1500 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1503 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1506 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1507 unsigned int order) { return NULL; }
1511 * Move the free pages in a range to the free lists of the requested type.
1512 * Note that start_page and end_pages are not aligned on a pageblock
1513 * boundary. If alignment is required, use move_freepages_block()
1515 int move_freepages(struct zone *zone,
1516 struct page *start_page, struct page *end_page,
1521 int pages_moved = 0;
1523 #ifndef CONFIG_HOLES_IN_ZONE
1525 * page_zone is not safe to call in this context when
1526 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1527 * anyway as we check zone boundaries in move_freepages_block().
1528 * Remove at a later date when no bug reports exist related to
1529 * grouping pages by mobility
1531 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1534 for (page = start_page; page <= end_page;) {
1535 /* Make sure we are not inadvertently changing nodes */
1536 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1538 if (!pfn_valid_within(page_to_pfn(page))) {
1543 if (!PageBuddy(page)) {
1548 order = page_order(page);
1549 list_move(&page->lru,
1550 &zone->free_area[order].free_list[migratetype]);
1552 pages_moved += 1 << order;
1558 int move_freepages_block(struct zone *zone, struct page *page,
1561 unsigned long start_pfn, end_pfn;
1562 struct page *start_page, *end_page;
1564 start_pfn = page_to_pfn(page);
1565 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1566 start_page = pfn_to_page(start_pfn);
1567 end_page = start_page + pageblock_nr_pages - 1;
1568 end_pfn = start_pfn + pageblock_nr_pages - 1;
1570 /* Do not cross zone boundaries */
1571 if (!zone_spans_pfn(zone, start_pfn))
1573 if (!zone_spans_pfn(zone, end_pfn))
1576 return move_freepages(zone, start_page, end_page, migratetype);
1579 static void change_pageblock_range(struct page *pageblock_page,
1580 int start_order, int migratetype)
1582 int nr_pageblocks = 1 << (start_order - pageblock_order);
1584 while (nr_pageblocks--) {
1585 set_pageblock_migratetype(pageblock_page, migratetype);
1586 pageblock_page += pageblock_nr_pages;
1591 * When we are falling back to another migratetype during allocation, try to
1592 * steal extra free pages from the same pageblocks to satisfy further
1593 * allocations, instead of polluting multiple pageblocks.
1595 * If we are stealing a relatively large buddy page, it is likely there will
1596 * be more free pages in the pageblock, so try to steal them all. For
1597 * reclaimable and unmovable allocations, we steal regardless of page size,
1598 * as fragmentation caused by those allocations polluting movable pageblocks
1599 * is worse than movable allocations stealing from unmovable and reclaimable
1602 static bool can_steal_fallback(unsigned int order, int start_mt)
1605 * Leaving this order check is intended, although there is
1606 * relaxed order check in next check. The reason is that
1607 * we can actually steal whole pageblock if this condition met,
1608 * but, below check doesn't guarantee it and that is just heuristic
1609 * so could be changed anytime.
1611 if (order >= pageblock_order)
1614 if (order >= pageblock_order / 2 ||
1615 start_mt == MIGRATE_RECLAIMABLE ||
1616 start_mt == MIGRATE_UNMOVABLE ||
1617 page_group_by_mobility_disabled)
1624 * This function implements actual steal behaviour. If order is large enough,
1625 * we can steal whole pageblock. If not, we first move freepages in this
1626 * pageblock and check whether half of pages are moved or not. If half of
1627 * pages are moved, we can change migratetype of pageblock and permanently
1628 * use it's pages as requested migratetype in the future.
1630 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1633 unsigned int current_order = page_order(page);
1636 /* Take ownership for orders >= pageblock_order */
1637 if (current_order >= pageblock_order) {
1638 change_pageblock_range(page, current_order, start_type);
1642 pages = move_freepages_block(zone, page, start_type);
1644 /* Claim the whole block if over half of it is free */
1645 if (pages >= (1 << (pageblock_order-1)) ||
1646 page_group_by_mobility_disabled)
1647 set_pageblock_migratetype(page, start_type);
1651 * Check whether there is a suitable fallback freepage with requested order.
1652 * If only_stealable is true, this function returns fallback_mt only if
1653 * we can steal other freepages all together. This would help to reduce
1654 * fragmentation due to mixed migratetype pages in one pageblock.
1656 int find_suitable_fallback(struct free_area *area, unsigned int order,
1657 int migratetype, bool only_stealable, bool *can_steal)
1662 if (area->nr_free == 0)
1667 fallback_mt = fallbacks[migratetype][i];
1668 if (fallback_mt == MIGRATE_TYPES)
1671 if (list_empty(&area->free_list[fallback_mt]))
1674 if (can_steal_fallback(order, migratetype))
1677 if (!only_stealable)
1688 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1689 * there are no empty page blocks that contain a page with a suitable order
1691 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1692 unsigned int alloc_order)
1695 unsigned long max_managed, flags;
1698 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1699 * Check is race-prone but harmless.
1701 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1702 if (zone->nr_reserved_highatomic >= max_managed)
1705 spin_lock_irqsave(&zone->lock, flags);
1707 /* Recheck the nr_reserved_highatomic limit under the lock */
1708 if (zone->nr_reserved_highatomic >= max_managed)
1712 mt = get_pageblock_migratetype(page);
1713 if (mt != MIGRATE_HIGHATOMIC &&
1714 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1715 zone->nr_reserved_highatomic += pageblock_nr_pages;
1716 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1717 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1721 spin_unlock_irqrestore(&zone->lock, flags);
1725 * Used when an allocation is about to fail under memory pressure. This
1726 * potentially hurts the reliability of high-order allocations when under
1727 * intense memory pressure but failed atomic allocations should be easier
1728 * to recover from than an OOM.
1730 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1732 struct zonelist *zonelist = ac->zonelist;
1733 unsigned long flags;
1739 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1741 /* Preserve at least one pageblock */
1742 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1745 spin_lock_irqsave(&zone->lock, flags);
1746 for (order = 0; order < MAX_ORDER; order++) {
1747 struct free_area *area = &(zone->free_area[order]);
1749 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1752 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1756 * It should never happen but changes to locking could
1757 * inadvertently allow a per-cpu drain to add pages
1758 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1759 * and watch for underflows.
1761 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1762 zone->nr_reserved_highatomic);
1765 * Convert to ac->migratetype and avoid the normal
1766 * pageblock stealing heuristics. Minimally, the caller
1767 * is doing the work and needs the pages. More
1768 * importantly, if the block was always converted to
1769 * MIGRATE_UNMOVABLE or another type then the number
1770 * of pageblocks that cannot be completely freed
1773 set_pageblock_migratetype(page, ac->migratetype);
1774 move_freepages_block(zone, page, ac->migratetype);
1775 spin_unlock_irqrestore(&zone->lock, flags);
1778 spin_unlock_irqrestore(&zone->lock, flags);
1782 /* Remove an element from the buddy allocator from the fallback list */
1783 static inline struct page *
1784 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1786 struct free_area *area;
1787 unsigned int current_order;
1792 /* Find the largest possible block of pages in the other list */
1793 for (current_order = MAX_ORDER-1;
1794 current_order >= order && current_order <= MAX_ORDER-1;
1796 area = &(zone->free_area[current_order]);
1797 fallback_mt = find_suitable_fallback(area, current_order,
1798 start_migratetype, false, &can_steal);
1799 if (fallback_mt == -1)
1802 page = list_entry(area->free_list[fallback_mt].next,
1805 steal_suitable_fallback(zone, page, start_migratetype);
1807 /* Remove the page from the freelists */
1809 list_del(&page->lru);
1810 rmv_page_order(page);
1812 expand(zone, page, order, current_order, area,
1815 * The pcppage_migratetype may differ from pageblock's
1816 * migratetype depending on the decisions in
1817 * find_suitable_fallback(). This is OK as long as it does not
1818 * differ for MIGRATE_CMA pageblocks. Those can be used as
1819 * fallback only via special __rmqueue_cma_fallback() function
1821 set_pcppage_migratetype(page, start_migratetype);
1823 trace_mm_page_alloc_extfrag(page, order, current_order,
1824 start_migratetype, fallback_mt);
1833 * Do the hard work of removing an element from the buddy allocator.
1834 * Call me with the zone->lock already held.
1836 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1837 int migratetype, gfp_t gfp_flags)
1841 page = __rmqueue_smallest(zone, order, migratetype);
1842 if (unlikely(!page)) {
1843 if (migratetype == MIGRATE_MOVABLE)
1844 page = __rmqueue_cma_fallback(zone, order);
1847 page = __rmqueue_fallback(zone, order, migratetype);
1850 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1855 * Obtain a specified number of elements from the buddy allocator, all under
1856 * a single hold of the lock, for efficiency. Add them to the supplied list.
1857 * Returns the number of new pages which were placed at *list.
1859 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1860 unsigned long count, struct list_head *list,
1861 int migratetype, bool cold)
1865 spin_lock(&zone->lock);
1866 for (i = 0; i < count; ++i) {
1867 struct page *page = __rmqueue(zone, order, migratetype, 0);
1868 if (unlikely(page == NULL))
1872 * Split buddy pages returned by expand() are received here
1873 * in physical page order. The page is added to the callers and
1874 * list and the list head then moves forward. From the callers
1875 * perspective, the linked list is ordered by page number in
1876 * some conditions. This is useful for IO devices that can
1877 * merge IO requests if the physical pages are ordered
1881 list_add(&page->lru, list);
1883 list_add_tail(&page->lru, list);
1885 if (is_migrate_cma(get_pcppage_migratetype(page)))
1886 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1889 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1890 spin_unlock(&zone->lock);
1896 * Called from the vmstat counter updater to drain pagesets of this
1897 * currently executing processor on remote nodes after they have
1900 * Note that this function must be called with the thread pinned to
1901 * a single processor.
1903 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1905 unsigned long flags;
1906 int to_drain, batch;
1908 local_irq_save(flags);
1909 batch = READ_ONCE(pcp->batch);
1910 to_drain = min(pcp->count, batch);
1912 free_pcppages_bulk(zone, to_drain, pcp);
1913 pcp->count -= to_drain;
1915 local_irq_restore(flags);
1920 * Drain pcplists of the indicated processor and zone.
1922 * The processor must either be the current processor and the
1923 * thread pinned to the current processor or a processor that
1926 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1928 unsigned long flags;
1929 struct per_cpu_pageset *pset;
1930 struct per_cpu_pages *pcp;
1932 local_irq_save(flags);
1933 pset = per_cpu_ptr(zone->pageset, cpu);
1937 free_pcppages_bulk(zone, pcp->count, pcp);
1940 local_irq_restore(flags);
1944 * Drain pcplists of all zones on the indicated processor.
1946 * The processor must either be the current processor and the
1947 * thread pinned to the current processor or a processor that
1950 static void drain_pages(unsigned int cpu)
1954 for_each_populated_zone(zone) {
1955 drain_pages_zone(cpu, zone);
1960 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1962 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1963 * the single zone's pages.
1965 void drain_local_pages(struct zone *zone)
1967 int cpu = smp_processor_id();
1970 drain_pages_zone(cpu, zone);
1976 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1978 * When zone parameter is non-NULL, spill just the single zone's pages.
1980 * Note that this code is protected against sending an IPI to an offline
1981 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1982 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1983 * nothing keeps CPUs from showing up after we populated the cpumask and
1984 * before the call to on_each_cpu_mask().
1986 void drain_all_pages(struct zone *zone)
1991 * Allocate in the BSS so we wont require allocation in
1992 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1994 static cpumask_t cpus_with_pcps;
1997 * We don't care about racing with CPU hotplug event
1998 * as offline notification will cause the notified
1999 * cpu to drain that CPU pcps and on_each_cpu_mask
2000 * disables preemption as part of its processing
2002 for_each_online_cpu(cpu) {
2003 struct per_cpu_pageset *pcp;
2005 bool has_pcps = false;
2008 pcp = per_cpu_ptr(zone->pageset, cpu);
2012 for_each_populated_zone(z) {
2013 pcp = per_cpu_ptr(z->pageset, cpu);
2014 if (pcp->pcp.count) {
2022 cpumask_set_cpu(cpu, &cpus_with_pcps);
2024 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2026 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2030 #ifdef CONFIG_HIBERNATION
2032 void mark_free_pages(struct zone *zone)
2034 unsigned long pfn, max_zone_pfn;
2035 unsigned long flags;
2036 unsigned int order, t;
2037 struct list_head *curr;
2039 if (zone_is_empty(zone))
2042 spin_lock_irqsave(&zone->lock, flags);
2044 max_zone_pfn = zone_end_pfn(zone);
2045 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2046 if (pfn_valid(pfn)) {
2047 struct page *page = pfn_to_page(pfn);
2049 if (!swsusp_page_is_forbidden(page))
2050 swsusp_unset_page_free(page);
2053 for_each_migratetype_order(order, t) {
2054 list_for_each(curr, &zone->free_area[order].free_list[t]) {
2057 pfn = page_to_pfn(list_entry(curr, struct page, lru));
2058 for (i = 0; i < (1UL << order); i++)
2059 swsusp_set_page_free(pfn_to_page(pfn + i));
2062 spin_unlock_irqrestore(&zone->lock, flags);
2064 #endif /* CONFIG_PM */
2067 * Free a 0-order page
2068 * cold == true ? free a cold page : free a hot page
2070 void free_hot_cold_page(struct page *page, bool cold)
2072 struct zone *zone = page_zone(page);
2073 struct per_cpu_pages *pcp;
2074 unsigned long flags;
2075 unsigned long pfn = page_to_pfn(page);
2078 if (!free_pages_prepare(page, 0))
2081 migratetype = get_pfnblock_migratetype(page, pfn);
2082 set_pcppage_migratetype(page, migratetype);
2083 local_irq_save(flags);
2084 __count_vm_event(PGFREE);
2087 * We only track unmovable, reclaimable and movable on pcp lists.
2088 * Free ISOLATE pages back to the allocator because they are being
2089 * offlined but treat RESERVE as movable pages so we can get those
2090 * areas back if necessary. Otherwise, we may have to free
2091 * excessively into the page allocator
2093 if (migratetype >= MIGRATE_PCPTYPES) {
2094 if (unlikely(is_migrate_isolate(migratetype))) {
2095 free_one_page(zone, page, pfn, 0, migratetype);
2098 migratetype = MIGRATE_MOVABLE;
2101 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2103 list_add(&page->lru, &pcp->lists[migratetype]);
2105 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2107 if (pcp->count >= pcp->high) {
2108 unsigned long batch = READ_ONCE(pcp->batch);
2109 free_pcppages_bulk(zone, batch, pcp);
2110 pcp->count -= batch;
2114 local_irq_restore(flags);
2118 * Free a list of 0-order pages
2120 void free_hot_cold_page_list(struct list_head *list, bool cold)
2122 struct page *page, *next;
2124 list_for_each_entry_safe(page, next, list, lru) {
2125 trace_mm_page_free_batched(page, cold);
2126 free_hot_cold_page(page, cold);
2131 * split_page takes a non-compound higher-order page, and splits it into
2132 * n (1<<order) sub-pages: page[0..n]
2133 * Each sub-page must be freed individually.
2135 * Note: this is probably too low level an operation for use in drivers.
2136 * Please consult with lkml before using this in your driver.
2138 void split_page(struct page *page, unsigned int order)
2143 VM_BUG_ON_PAGE(PageCompound(page), page);
2144 VM_BUG_ON_PAGE(!page_count(page), page);
2146 #ifdef CONFIG_KMEMCHECK
2148 * Split shadow pages too, because free(page[0]) would
2149 * otherwise free the whole shadow.
2151 if (kmemcheck_page_is_tracked(page))
2152 split_page(virt_to_page(page[0].shadow), order);
2155 gfp_mask = get_page_owner_gfp(page);
2156 set_page_owner(page, 0, gfp_mask);
2157 for (i = 1; i < (1 << order); i++) {
2158 set_page_refcounted(page + i);
2159 set_page_owner(page + i, 0, gfp_mask);
2162 EXPORT_SYMBOL_GPL(split_page);
2164 int __isolate_free_page(struct page *page, unsigned int order)
2166 unsigned long watermark;
2170 BUG_ON(!PageBuddy(page));
2172 zone = page_zone(page);
2173 mt = get_pageblock_migratetype(page);
2175 if (!is_migrate_isolate(mt)) {
2176 /* Obey watermarks as if the page was being allocated */
2177 watermark = low_wmark_pages(zone) + (1 << order);
2178 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2181 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2184 /* Remove page from free list */
2185 list_del(&page->lru);
2186 zone->free_area[order].nr_free--;
2187 rmv_page_order(page);
2189 set_page_owner(page, order, __GFP_MOVABLE);
2191 /* Set the pageblock if the isolated page is at least a pageblock */
2192 if (order >= pageblock_order - 1) {
2193 struct page *endpage = page + (1 << order) - 1;
2194 for (; page < endpage; page += pageblock_nr_pages) {
2195 int mt = get_pageblock_migratetype(page);
2196 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2197 set_pageblock_migratetype(page,
2203 return 1UL << order;
2207 * Similar to split_page except the page is already free. As this is only
2208 * being used for migration, the migratetype of the block also changes.
2209 * As this is called with interrupts disabled, the caller is responsible
2210 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2213 * Note: this is probably too low level an operation for use in drivers.
2214 * Please consult with lkml before using this in your driver.
2216 int split_free_page(struct page *page)
2221 order = page_order(page);
2223 nr_pages = __isolate_free_page(page, order);
2227 /* Split into individual pages */
2228 set_page_refcounted(page);
2229 split_page(page, order);
2234 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2237 struct page *buffered_rmqueue(struct zone *preferred_zone,
2238 struct zone *zone, unsigned int order,
2239 gfp_t gfp_flags, int alloc_flags, int migratetype)
2241 unsigned long flags;
2243 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2245 if (likely(order == 0)) {
2246 struct per_cpu_pages *pcp;
2247 struct list_head *list;
2249 local_irq_save(flags);
2250 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2251 list = &pcp->lists[migratetype];
2252 if (list_empty(list)) {
2253 pcp->count += rmqueue_bulk(zone, 0,
2256 if (unlikely(list_empty(list)))
2261 page = list_entry(list->prev, struct page, lru);
2263 page = list_entry(list->next, struct page, lru);
2265 list_del(&page->lru);
2268 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2270 * __GFP_NOFAIL is not to be used in new code.
2272 * All __GFP_NOFAIL callers should be fixed so that they
2273 * properly detect and handle allocation failures.
2275 * We most definitely don't want callers attempting to
2276 * allocate greater than order-1 page units with
2279 WARN_ON_ONCE(order > 1);
2281 spin_lock_irqsave(&zone->lock, flags);
2284 if (alloc_flags & ALLOC_HARDER) {
2285 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2287 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2290 page = __rmqueue(zone, order, migratetype, gfp_flags);
2291 spin_unlock(&zone->lock);
2294 __mod_zone_freepage_state(zone, -(1 << order),
2295 get_pcppage_migratetype(page));
2298 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2299 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2300 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2301 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2303 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2304 zone_statistics(preferred_zone, zone, gfp_flags);
2305 local_irq_restore(flags);
2307 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2311 local_irq_restore(flags);
2315 #ifdef CONFIG_FAIL_PAGE_ALLOC
2318 struct fault_attr attr;
2320 bool ignore_gfp_highmem;
2321 bool ignore_gfp_reclaim;
2323 } fail_page_alloc = {
2324 .attr = FAULT_ATTR_INITIALIZER,
2325 .ignore_gfp_reclaim = true,
2326 .ignore_gfp_highmem = true,
2330 static int __init setup_fail_page_alloc(char *str)
2332 return setup_fault_attr(&fail_page_alloc.attr, str);
2334 __setup("fail_page_alloc=", setup_fail_page_alloc);
2336 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2338 if (order < fail_page_alloc.min_order)
2340 if (gfp_mask & __GFP_NOFAIL)
2342 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2344 if (fail_page_alloc.ignore_gfp_reclaim &&
2345 (gfp_mask & __GFP_DIRECT_RECLAIM))
2348 return should_fail(&fail_page_alloc.attr, 1 << order);
2351 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2353 static int __init fail_page_alloc_debugfs(void)
2355 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2358 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2359 &fail_page_alloc.attr);
2361 return PTR_ERR(dir);
2363 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2364 &fail_page_alloc.ignore_gfp_reclaim))
2366 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2367 &fail_page_alloc.ignore_gfp_highmem))
2369 if (!debugfs_create_u32("min-order", mode, dir,
2370 &fail_page_alloc.min_order))
2375 debugfs_remove_recursive(dir);
2380 late_initcall(fail_page_alloc_debugfs);
2382 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2384 #else /* CONFIG_FAIL_PAGE_ALLOC */
2386 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2391 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2394 * Return true if free base pages are above 'mark'. For high-order checks it
2395 * will return true of the order-0 watermark is reached and there is at least
2396 * one free page of a suitable size. Checking now avoids taking the zone lock
2397 * to check in the allocation paths if no pages are free.
2399 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2400 unsigned long mark, int classzone_idx, int alloc_flags,
2405 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2407 /* free_pages may go negative - that's OK */
2408 free_pages -= (1 << order) - 1;
2410 if (alloc_flags & ALLOC_HIGH)
2414 * If the caller does not have rights to ALLOC_HARDER then subtract
2415 * the high-atomic reserves. This will over-estimate the size of the
2416 * atomic reserve but it avoids a search.
2418 if (likely(!alloc_harder))
2419 free_pages -= z->nr_reserved_highatomic;
2424 /* If allocation can't use CMA areas don't use free CMA pages */
2425 if (!(alloc_flags & ALLOC_CMA))
2426 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2430 * Check watermarks for an order-0 allocation request. If these
2431 * are not met, then a high-order request also cannot go ahead
2432 * even if a suitable page happened to be free.
2434 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2437 /* If this is an order-0 request then the watermark is fine */
2441 /* For a high-order request, check at least one suitable page is free */
2442 for (o = order; o < MAX_ORDER; o++) {
2443 struct free_area *area = &z->free_area[o];
2452 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2453 if (!list_empty(&area->free_list[mt]))
2458 if ((alloc_flags & ALLOC_CMA) &&
2459 !list_empty(&area->free_list[MIGRATE_CMA])) {
2467 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2468 int classzone_idx, int alloc_flags)
2470 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2471 zone_page_state(z, NR_FREE_PAGES));
2474 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2475 unsigned long mark, int classzone_idx)
2477 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2479 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2480 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2482 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2487 static bool zone_local(struct zone *local_zone, struct zone *zone)
2489 return local_zone->node == zone->node;
2492 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2494 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
2497 #else /* CONFIG_NUMA */
2498 static bool zone_local(struct zone *local_zone, struct zone *zone)
2503 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2507 #endif /* CONFIG_NUMA */
2509 static void reset_alloc_batches(struct zone *preferred_zone)
2511 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2514 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2515 high_wmark_pages(zone) - low_wmark_pages(zone) -
2516 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2517 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2518 } while (zone++ != preferred_zone);
2522 * get_page_from_freelist goes through the zonelist trying to allocate
2525 static struct page *
2526 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2527 const struct alloc_context *ac)
2529 struct zonelist *zonelist = ac->zonelist;
2531 struct page *page = NULL;
2533 int nr_fair_skipped = 0;
2534 bool zonelist_rescan;
2537 zonelist_rescan = false;
2540 * Scan zonelist, looking for a zone with enough free.
2541 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2543 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2547 if (cpusets_enabled() &&
2548 (alloc_flags & ALLOC_CPUSET) &&
2549 !cpuset_zone_allowed(zone, gfp_mask))
2552 * Distribute pages in proportion to the individual
2553 * zone size to ensure fair page aging. The zone a
2554 * page was allocated in should have no effect on the
2555 * time the page has in memory before being reclaimed.
2557 if (alloc_flags & ALLOC_FAIR) {
2558 if (!zone_local(ac->preferred_zone, zone))
2560 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2566 * When allocating a page cache page for writing, we
2567 * want to get it from a zone that is within its dirty
2568 * limit, such that no single zone holds more than its
2569 * proportional share of globally allowed dirty pages.
2570 * The dirty limits take into account the zone's
2571 * lowmem reserves and high watermark so that kswapd
2572 * should be able to balance it without having to
2573 * write pages from its LRU list.
2575 * This may look like it could increase pressure on
2576 * lower zones by failing allocations in higher zones
2577 * before they are full. But the pages that do spill
2578 * over are limited as the lower zones are protected
2579 * by this very same mechanism. It should not become
2580 * a practical burden to them.
2582 * XXX: For now, allow allocations to potentially
2583 * exceed the per-zone dirty limit in the slowpath
2584 * (spread_dirty_pages unset) before going into reclaim,
2585 * which is important when on a NUMA setup the allowed
2586 * zones are together not big enough to reach the
2587 * global limit. The proper fix for these situations
2588 * will require awareness of zones in the
2589 * dirty-throttling and the flusher threads.
2591 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2594 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2595 if (!zone_watermark_ok(zone, order, mark,
2596 ac->classzone_idx, alloc_flags)) {
2599 /* Checked here to keep the fast path fast */
2600 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2601 if (alloc_flags & ALLOC_NO_WATERMARKS)
2604 if (zone_reclaim_mode == 0 ||
2605 !zone_allows_reclaim(ac->preferred_zone, zone))
2608 ret = zone_reclaim(zone, gfp_mask, order);
2610 case ZONE_RECLAIM_NOSCAN:
2613 case ZONE_RECLAIM_FULL:
2614 /* scanned but unreclaimable */
2617 /* did we reclaim enough */
2618 if (zone_watermark_ok(zone, order, mark,
2619 ac->classzone_idx, alloc_flags))
2627 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2628 gfp_mask, alloc_flags, ac->migratetype);
2630 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2634 * If this is a high-order atomic allocation then check
2635 * if the pageblock should be reserved for the future
2637 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2638 reserve_highatomic_pageblock(page, zone, order);
2645 * The first pass makes sure allocations are spread fairly within the
2646 * local node. However, the local node might have free pages left
2647 * after the fairness batches are exhausted, and remote zones haven't
2648 * even been considered yet. Try once more without fairness, and
2649 * include remote zones now, before entering the slowpath and waking
2650 * kswapd: prefer spilling to a remote zone over swapping locally.
2652 if (alloc_flags & ALLOC_FAIR) {
2653 alloc_flags &= ~ALLOC_FAIR;
2654 if (nr_fair_skipped) {
2655 zonelist_rescan = true;
2656 reset_alloc_batches(ac->preferred_zone);
2658 if (nr_online_nodes > 1)
2659 zonelist_rescan = true;
2662 if (zonelist_rescan)
2669 * Large machines with many possible nodes should not always dump per-node
2670 * meminfo in irq context.
2672 static inline bool should_suppress_show_mem(void)
2677 ret = in_interrupt();
2682 static DEFINE_RATELIMIT_STATE(nopage_rs,
2683 DEFAULT_RATELIMIT_INTERVAL,
2684 DEFAULT_RATELIMIT_BURST);
2686 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2688 unsigned int filter = SHOW_MEM_FILTER_NODES;
2690 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2691 debug_guardpage_minorder() > 0)
2695 * This documents exceptions given to allocations in certain
2696 * contexts that are allowed to allocate outside current's set
2699 if (!(gfp_mask & __GFP_NOMEMALLOC))
2700 if (test_thread_flag(TIF_MEMDIE) ||
2701 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2702 filter &= ~SHOW_MEM_FILTER_NODES;
2703 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2704 filter &= ~SHOW_MEM_FILTER_NODES;
2707 struct va_format vaf;
2710 va_start(args, fmt);
2715 pr_warn("%pV", &vaf);
2720 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2721 current->comm, order, gfp_mask);
2724 if (!should_suppress_show_mem())
2728 static inline struct page *
2729 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2730 const struct alloc_context *ac, unsigned long *did_some_progress)
2732 struct oom_control oc = {
2733 .zonelist = ac->zonelist,
2734 .nodemask = ac->nodemask,
2735 .gfp_mask = gfp_mask,
2740 *did_some_progress = 0;
2743 * Acquire the oom lock. If that fails, somebody else is
2744 * making progress for us.
2746 if (!mutex_trylock(&oom_lock)) {
2747 *did_some_progress = 1;
2748 schedule_timeout_uninterruptible(1);
2753 * Go through the zonelist yet one more time, keep very high watermark
2754 * here, this is only to catch a parallel oom killing, we must fail if
2755 * we're still under heavy pressure.
2757 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2758 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2762 if (!(gfp_mask & __GFP_NOFAIL)) {
2763 /* Coredumps can quickly deplete all memory reserves */
2764 if (current->flags & PF_DUMPCORE)
2766 /* The OOM killer will not help higher order allocs */
2767 if (order > PAGE_ALLOC_COSTLY_ORDER)
2769 /* The OOM killer does not needlessly kill tasks for lowmem */
2770 if (ac->high_zoneidx < ZONE_NORMAL)
2772 /* The OOM killer does not compensate for IO-less reclaim */
2773 if (!(gfp_mask & __GFP_FS)) {
2775 * XXX: Page reclaim didn't yield anything,
2776 * and the OOM killer can't be invoked, but
2777 * keep looping as per tradition.
2779 *did_some_progress = 1;
2782 if (pm_suspended_storage())
2784 /* The OOM killer may not free memory on a specific node */
2785 if (gfp_mask & __GFP_THISNODE)
2788 /* Exhausted what can be done so it's blamo time */
2789 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2790 *did_some_progress = 1;
2792 mutex_unlock(&oom_lock);
2796 #ifdef CONFIG_COMPACTION
2797 /* Try memory compaction for high-order allocations before reclaim */
2798 static struct page *
2799 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2800 int alloc_flags, const struct alloc_context *ac,
2801 enum migrate_mode mode, int *contended_compaction,
2802 bool *deferred_compaction)
2804 unsigned long compact_result;
2810 current->flags |= PF_MEMALLOC;
2811 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2812 mode, contended_compaction);
2813 current->flags &= ~PF_MEMALLOC;
2815 switch (compact_result) {
2816 case COMPACT_DEFERRED:
2817 *deferred_compaction = true;
2819 case COMPACT_SKIPPED:
2826 * At least in one zone compaction wasn't deferred or skipped, so let's
2827 * count a compaction stall
2829 count_vm_event(COMPACTSTALL);
2831 page = get_page_from_freelist(gfp_mask, order,
2832 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2835 struct zone *zone = page_zone(page);
2837 zone->compact_blockskip_flush = false;
2838 compaction_defer_reset(zone, order, true);
2839 count_vm_event(COMPACTSUCCESS);
2844 * It's bad if compaction run occurs and fails. The most likely reason
2845 * is that pages exist, but not enough to satisfy watermarks.
2847 count_vm_event(COMPACTFAIL);
2854 static inline struct page *
2855 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2856 int alloc_flags, const struct alloc_context *ac,
2857 enum migrate_mode mode, int *contended_compaction,
2858 bool *deferred_compaction)
2862 #endif /* CONFIG_COMPACTION */
2864 /* Perform direct synchronous page reclaim */
2866 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2867 const struct alloc_context *ac)
2869 struct reclaim_state reclaim_state;
2874 /* We now go into synchronous reclaim */
2875 cpuset_memory_pressure_bump();
2876 current->flags |= PF_MEMALLOC;
2877 lockdep_set_current_reclaim_state(gfp_mask);
2878 reclaim_state.reclaimed_slab = 0;
2879 current->reclaim_state = &reclaim_state;
2881 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2884 current->reclaim_state = NULL;
2885 lockdep_clear_current_reclaim_state();
2886 current->flags &= ~PF_MEMALLOC;
2893 /* The really slow allocator path where we enter direct reclaim */
2894 static inline struct page *
2895 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2896 int alloc_flags, const struct alloc_context *ac,
2897 unsigned long *did_some_progress)
2899 struct page *page = NULL;
2900 bool drained = false;
2902 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2903 if (unlikely(!(*did_some_progress)))
2907 page = get_page_from_freelist(gfp_mask, order,
2908 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2911 * If an allocation failed after direct reclaim, it could be because
2912 * pages are pinned on the per-cpu lists or in high alloc reserves.
2913 * Shrink them them and try again
2915 if (!page && !drained) {
2916 unreserve_highatomic_pageblock(ac);
2917 drain_all_pages(NULL);
2926 * This is called in the allocator slow-path if the allocation request is of
2927 * sufficient urgency to ignore watermarks and take other desperate measures
2929 static inline struct page *
2930 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2931 const struct alloc_context *ac)
2936 page = get_page_from_freelist(gfp_mask, order,
2937 ALLOC_NO_WATERMARKS, ac);
2939 if (!page && gfp_mask & __GFP_NOFAIL)
2940 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2942 } while (!page && (gfp_mask & __GFP_NOFAIL));
2947 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2952 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2953 ac->high_zoneidx, ac->nodemask)
2954 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2958 gfp_to_alloc_flags(gfp_t gfp_mask)
2960 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2962 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2963 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2966 * The caller may dip into page reserves a bit more if the caller
2967 * cannot run direct reclaim, or if the caller has realtime scheduling
2968 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2969 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2971 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2973 if (gfp_mask & __GFP_ATOMIC) {
2975 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2976 * if it can't schedule.
2978 if (!(gfp_mask & __GFP_NOMEMALLOC))
2979 alloc_flags |= ALLOC_HARDER;
2981 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2982 * comment for __cpuset_node_allowed().
2984 alloc_flags &= ~ALLOC_CPUSET;
2985 } else if (unlikely(rt_task(current)) && !in_interrupt())
2986 alloc_flags |= ALLOC_HARDER;
2988 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2989 if (gfp_mask & __GFP_MEMALLOC)
2990 alloc_flags |= ALLOC_NO_WATERMARKS;
2991 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2992 alloc_flags |= ALLOC_NO_WATERMARKS;
2993 else if (!in_interrupt() &&
2994 ((current->flags & PF_MEMALLOC) ||
2995 unlikely(test_thread_flag(TIF_MEMDIE))))
2996 alloc_flags |= ALLOC_NO_WATERMARKS;
2999 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3000 alloc_flags |= ALLOC_CMA;
3005 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3007 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3010 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3012 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3015 static inline struct page *
3016 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3017 struct alloc_context *ac)
3019 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3020 struct page *page = NULL;
3022 unsigned long pages_reclaimed = 0;
3023 unsigned long did_some_progress;
3024 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3025 bool deferred_compaction = false;
3026 int contended_compaction = COMPACT_CONTENDED_NONE;
3029 * In the slowpath, we sanity check order to avoid ever trying to
3030 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3031 * be using allocators in order of preference for an area that is
3034 if (order >= MAX_ORDER) {
3035 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3040 * We also sanity check to catch abuse of atomic reserves being used by
3041 * callers that are not in atomic context.
3043 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3044 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3045 gfp_mask &= ~__GFP_ATOMIC;
3048 * If this allocation cannot block and it is for a specific node, then
3049 * fail early. There's no need to wakeup kswapd or retry for a
3050 * speculative node-specific allocation.
3052 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3056 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3057 wake_all_kswapds(order, ac);
3060 * OK, we're below the kswapd watermark and have kicked background
3061 * reclaim. Now things get more complex, so set up alloc_flags according
3062 * to how we want to proceed.
3064 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3067 * Find the true preferred zone if the allocation is unconstrained by
3070 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3071 struct zoneref *preferred_zoneref;
3072 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3073 ac->high_zoneidx, NULL, &ac->preferred_zone);
3074 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3077 /* This is the last chance, in general, before the goto nopage. */
3078 page = get_page_from_freelist(gfp_mask, order,
3079 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3083 /* Allocate without watermarks if the context allows */
3084 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3086 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3087 * the allocation is high priority and these type of
3088 * allocations are system rather than user orientated
3090 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3092 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3099 /* Caller is not willing to reclaim, we can't balance anything */
3100 if (!can_direct_reclaim) {
3102 * All existing users of the deprecated __GFP_NOFAIL are
3103 * blockable, so warn of any new users that actually allow this
3104 * type of allocation to fail.
3106 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3110 /* Avoid recursion of direct reclaim */
3111 if (current->flags & PF_MEMALLOC)
3114 /* Avoid allocations with no watermarks from looping endlessly */
3115 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3119 * Try direct compaction. The first pass is asynchronous. Subsequent
3120 * attempts after direct reclaim are synchronous
3122 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3124 &contended_compaction,
3125 &deferred_compaction);
3129 /* Checks for THP-specific high-order allocations */
3130 if (is_thp_gfp_mask(gfp_mask)) {
3132 * If compaction is deferred for high-order allocations, it is
3133 * because sync compaction recently failed. If this is the case
3134 * and the caller requested a THP allocation, we do not want
3135 * to heavily disrupt the system, so we fail the allocation
3136 * instead of entering direct reclaim.
3138 if (deferred_compaction)
3142 * In all zones where compaction was attempted (and not
3143 * deferred or skipped), lock contention has been detected.
3144 * For THP allocation we do not want to disrupt the others
3145 * so we fallback to base pages instead.
3147 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3151 * If compaction was aborted due to need_resched(), we do not
3152 * want to further increase allocation latency, unless it is
3153 * khugepaged trying to collapse.
3155 if (contended_compaction == COMPACT_CONTENDED_SCHED
3156 && !(current->flags & PF_KTHREAD))
3161 * It can become very expensive to allocate transparent hugepages at
3162 * fault, so use asynchronous memory compaction for THP unless it is
3163 * khugepaged trying to collapse.
3165 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3166 migration_mode = MIGRATE_SYNC_LIGHT;
3168 /* Try direct reclaim and then allocating */
3169 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3170 &did_some_progress);
3174 /* Do not loop if specifically requested */
3175 if (gfp_mask & __GFP_NORETRY)
3178 /* Keep reclaiming pages as long as there is reasonable progress */
3179 pages_reclaimed += did_some_progress;
3180 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3181 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3182 /* Wait for some write requests to complete then retry */
3183 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3187 /* Reclaim has failed us, start killing things */
3188 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3192 /* Retry as long as the OOM killer is making progress */
3193 if (did_some_progress)
3198 * High-order allocations do not necessarily loop after
3199 * direct reclaim and reclaim/compaction depends on compaction
3200 * being called after reclaim so call directly if necessary
3202 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3204 &contended_compaction,
3205 &deferred_compaction);
3209 warn_alloc_failed(gfp_mask, order, NULL);
3215 * This is the 'heart' of the zoned buddy allocator.
3218 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3219 struct zonelist *zonelist, nodemask_t *nodemask)
3221 struct zoneref *preferred_zoneref;
3222 struct page *page = NULL;
3223 unsigned int cpuset_mems_cookie;
3224 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3225 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3226 struct alloc_context ac = {
3227 .high_zoneidx = gfp_zone(gfp_mask),
3228 .nodemask = nodemask,
3229 .migratetype = gfpflags_to_migratetype(gfp_mask),
3232 gfp_mask &= gfp_allowed_mask;
3234 lockdep_trace_alloc(gfp_mask);
3236 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3238 if (should_fail_alloc_page(gfp_mask, order))
3242 * Check the zones suitable for the gfp_mask contain at least one
3243 * valid zone. It's possible to have an empty zonelist as a result
3244 * of __GFP_THISNODE and a memoryless node
3246 if (unlikely(!zonelist->_zonerefs->zone))
3249 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3250 alloc_flags |= ALLOC_CMA;
3253 cpuset_mems_cookie = read_mems_allowed_begin();
3255 /* We set it here, as __alloc_pages_slowpath might have changed it */
3256 ac.zonelist = zonelist;
3258 /* Dirty zone balancing only done in the fast path */
3259 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3261 /* The preferred zone is used for statistics later */
3262 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3263 ac.nodemask ? : &cpuset_current_mems_allowed,
3264 &ac.preferred_zone);
3265 if (!ac.preferred_zone)
3267 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3269 /* First allocation attempt */
3270 alloc_mask = gfp_mask|__GFP_HARDWALL;
3271 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3272 if (unlikely(!page)) {
3274 * Runtime PM, block IO and its error handling path
3275 * can deadlock because I/O on the device might not
3278 alloc_mask = memalloc_noio_flags(gfp_mask);
3279 ac.spread_dirty_pages = false;
3281 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3284 if (kmemcheck_enabled && page)
3285 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3287 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3291 * When updating a task's mems_allowed, it is possible to race with
3292 * parallel threads in such a way that an allocation can fail while
3293 * the mask is being updated. If a page allocation is about to fail,
3294 * check if the cpuset changed during allocation and if so, retry.
3296 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3301 EXPORT_SYMBOL(__alloc_pages_nodemask);
3304 * Common helper functions.
3306 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3311 * __get_free_pages() returns a 32-bit address, which cannot represent
3314 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3316 page = alloc_pages(gfp_mask, order);
3319 return (unsigned long) page_address(page);
3321 EXPORT_SYMBOL(__get_free_pages);
3323 unsigned long get_zeroed_page(gfp_t gfp_mask)
3325 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3327 EXPORT_SYMBOL(get_zeroed_page);
3329 void __free_pages(struct page *page, unsigned int order)
3331 if (put_page_testzero(page)) {
3333 free_hot_cold_page(page, false);
3335 __free_pages_ok(page, order);
3339 EXPORT_SYMBOL(__free_pages);
3341 void free_pages(unsigned long addr, unsigned int order)
3344 VM_BUG_ON(!virt_addr_valid((void *)addr));
3345 __free_pages(virt_to_page((void *)addr), order);
3349 EXPORT_SYMBOL(free_pages);
3353 * An arbitrary-length arbitrary-offset area of memory which resides
3354 * within a 0 or higher order page. Multiple fragments within that page
3355 * are individually refcounted, in the page's reference counter.
3357 * The page_frag functions below provide a simple allocation framework for
3358 * page fragments. This is used by the network stack and network device
3359 * drivers to provide a backing region of memory for use as either an
3360 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3362 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3365 struct page *page = NULL;
3366 gfp_t gfp = gfp_mask;
3368 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3369 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3371 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3372 PAGE_FRAG_CACHE_MAX_ORDER);
3373 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3375 if (unlikely(!page))
3376 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3378 nc->va = page ? page_address(page) : NULL;
3383 void *__alloc_page_frag(struct page_frag_cache *nc,
3384 unsigned int fragsz, gfp_t gfp_mask)
3386 unsigned int size = PAGE_SIZE;
3390 if (unlikely(!nc->va)) {
3392 page = __page_frag_refill(nc, gfp_mask);
3396 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3397 /* if size can vary use size else just use PAGE_SIZE */
3400 /* Even if we own the page, we do not use atomic_set().
3401 * This would break get_page_unless_zero() users.
3403 atomic_add(size - 1, &page->_count);
3405 /* reset page count bias and offset to start of new frag */
3406 nc->pfmemalloc = page_is_pfmemalloc(page);
3407 nc->pagecnt_bias = size;
3411 offset = nc->offset - fragsz;
3412 if (unlikely(offset < 0)) {
3413 page = virt_to_page(nc->va);
3415 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3418 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3419 /* if size can vary use size else just use PAGE_SIZE */
3422 /* OK, page count is 0, we can safely set it */
3423 atomic_set(&page->_count, size);
3425 /* reset page count bias and offset to start of new frag */
3426 nc->pagecnt_bias = size;
3427 offset = size - fragsz;
3431 nc->offset = offset;
3433 return nc->va + offset;
3435 EXPORT_SYMBOL(__alloc_page_frag);
3438 * Frees a page fragment allocated out of either a compound or order 0 page.
3440 void __free_page_frag(void *addr)
3442 struct page *page = virt_to_head_page(addr);
3444 if (unlikely(put_page_testzero(page)))
3445 __free_pages_ok(page, compound_order(page));
3447 EXPORT_SYMBOL(__free_page_frag);
3450 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3451 * of the current memory cgroup.
3453 * It should be used when the caller would like to use kmalloc, but since the
3454 * allocation is large, it has to fall back to the page allocator.
3456 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3460 page = alloc_pages(gfp_mask, order);
3461 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3462 __free_pages(page, order);
3468 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3472 page = alloc_pages_node(nid, gfp_mask, order);
3473 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3474 __free_pages(page, order);
3481 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3484 void __free_kmem_pages(struct page *page, unsigned int order)
3486 memcg_kmem_uncharge(page, order);
3487 __free_pages(page, order);
3490 void free_kmem_pages(unsigned long addr, unsigned int order)
3493 VM_BUG_ON(!virt_addr_valid((void *)addr));
3494 __free_kmem_pages(virt_to_page((void *)addr), order);
3498 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3502 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3503 unsigned long used = addr + PAGE_ALIGN(size);
3505 split_page(virt_to_page((void *)addr), order);
3506 while (used < alloc_end) {
3511 return (void *)addr;
3515 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3516 * @size: the number of bytes to allocate
3517 * @gfp_mask: GFP flags for the allocation
3519 * This function is similar to alloc_pages(), except that it allocates the
3520 * minimum number of pages to satisfy the request. alloc_pages() can only
3521 * allocate memory in power-of-two pages.
3523 * This function is also limited by MAX_ORDER.
3525 * Memory allocated by this function must be released by free_pages_exact().
3527 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3529 unsigned int order = get_order(size);
3532 addr = __get_free_pages(gfp_mask, order);
3533 return make_alloc_exact(addr, order, size);
3535 EXPORT_SYMBOL(alloc_pages_exact);
3538 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3540 * @nid: the preferred node ID where memory should be allocated
3541 * @size: the number of bytes to allocate
3542 * @gfp_mask: GFP flags for the allocation
3544 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3547 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3549 unsigned int order = get_order(size);
3550 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3553 return make_alloc_exact((unsigned long)page_address(p), order, size);
3557 * free_pages_exact - release memory allocated via alloc_pages_exact()
3558 * @virt: the value returned by alloc_pages_exact.
3559 * @size: size of allocation, same value as passed to alloc_pages_exact().
3561 * Release the memory allocated by a previous call to alloc_pages_exact.
3563 void free_pages_exact(void *virt, size_t size)
3565 unsigned long addr = (unsigned long)virt;
3566 unsigned long end = addr + PAGE_ALIGN(size);
3568 while (addr < end) {
3573 EXPORT_SYMBOL(free_pages_exact);
3576 * nr_free_zone_pages - count number of pages beyond high watermark
3577 * @offset: The zone index of the highest zone
3579 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3580 * high watermark within all zones at or below a given zone index. For each
3581 * zone, the number of pages is calculated as:
3582 * managed_pages - high_pages
3584 static unsigned long nr_free_zone_pages(int offset)
3589 /* Just pick one node, since fallback list is circular */
3590 unsigned long sum = 0;
3592 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3594 for_each_zone_zonelist(zone, z, zonelist, offset) {
3595 unsigned long size = zone->managed_pages;
3596 unsigned long high = high_wmark_pages(zone);
3605 * nr_free_buffer_pages - count number of pages beyond high watermark
3607 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3608 * watermark within ZONE_DMA and ZONE_NORMAL.
3610 unsigned long nr_free_buffer_pages(void)
3612 return nr_free_zone_pages(gfp_zone(GFP_USER));
3614 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3617 * nr_free_pagecache_pages - count number of pages beyond high watermark
3619 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3620 * high watermark within all zones.
3622 unsigned long nr_free_pagecache_pages(void)
3624 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3627 static inline void show_node(struct zone *zone)
3629 if (IS_ENABLED(CONFIG_NUMA))
3630 printk("Node %d ", zone_to_nid(zone));
3633 void si_meminfo(struct sysinfo *val)
3635 val->totalram = totalram_pages;
3636 val->sharedram = global_page_state(NR_SHMEM);
3637 val->freeram = global_page_state(NR_FREE_PAGES);
3638 val->bufferram = nr_blockdev_pages();
3639 val->totalhigh = totalhigh_pages;
3640 val->freehigh = nr_free_highpages();
3641 val->mem_unit = PAGE_SIZE;
3644 EXPORT_SYMBOL(si_meminfo);
3647 void si_meminfo_node(struct sysinfo *val, int nid)
3649 int zone_type; /* needs to be signed */
3650 unsigned long managed_pages = 0;
3651 pg_data_t *pgdat = NODE_DATA(nid);
3653 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3654 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3655 val->totalram = managed_pages;
3656 val->sharedram = node_page_state(nid, NR_SHMEM);
3657 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3658 #ifdef CONFIG_HIGHMEM
3659 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3660 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3666 val->mem_unit = PAGE_SIZE;
3671 * Determine whether the node should be displayed or not, depending on whether
3672 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3674 bool skip_free_areas_node(unsigned int flags, int nid)
3677 unsigned int cpuset_mems_cookie;
3679 if (!(flags & SHOW_MEM_FILTER_NODES))
3683 cpuset_mems_cookie = read_mems_allowed_begin();
3684 ret = !node_isset(nid, cpuset_current_mems_allowed);
3685 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3690 #define K(x) ((x) << (PAGE_SHIFT-10))
3692 static void show_migration_types(unsigned char type)
3694 static const char types[MIGRATE_TYPES] = {
3695 [MIGRATE_UNMOVABLE] = 'U',
3696 [MIGRATE_MOVABLE] = 'M',
3697 [MIGRATE_RECLAIMABLE] = 'E',
3698 [MIGRATE_HIGHATOMIC] = 'H',
3700 [MIGRATE_CMA] = 'C',
3702 #ifdef CONFIG_MEMORY_ISOLATION
3703 [MIGRATE_ISOLATE] = 'I',
3706 char tmp[MIGRATE_TYPES + 1];
3710 for (i = 0; i < MIGRATE_TYPES; i++) {
3711 if (type & (1 << i))
3716 printk("(%s) ", tmp);
3720 * Show free area list (used inside shift_scroll-lock stuff)
3721 * We also calculate the percentage fragmentation. We do this by counting the
3722 * memory on each free list with the exception of the first item on the list.
3725 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3728 void show_free_areas(unsigned int filter)
3730 unsigned long free_pcp = 0;
3734 for_each_populated_zone(zone) {
3735 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3738 for_each_online_cpu(cpu)
3739 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3742 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3743 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3744 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3745 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3746 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3747 " free:%lu free_pcp:%lu free_cma:%lu\n",
3748 global_page_state(NR_ACTIVE_ANON),
3749 global_page_state(NR_INACTIVE_ANON),
3750 global_page_state(NR_ISOLATED_ANON),
3751 global_page_state(NR_ACTIVE_FILE),
3752 global_page_state(NR_INACTIVE_FILE),
3753 global_page_state(NR_ISOLATED_FILE),
3754 global_page_state(NR_UNEVICTABLE),
3755 global_page_state(NR_FILE_DIRTY),
3756 global_page_state(NR_WRITEBACK),
3757 global_page_state(NR_UNSTABLE_NFS),
3758 global_page_state(NR_SLAB_RECLAIMABLE),
3759 global_page_state(NR_SLAB_UNRECLAIMABLE),
3760 global_page_state(NR_FILE_MAPPED),
3761 global_page_state(NR_SHMEM),
3762 global_page_state(NR_PAGETABLE),
3763 global_page_state(NR_BOUNCE),
3764 global_page_state(NR_FREE_PAGES),
3766 global_page_state(NR_FREE_CMA_PAGES));
3768 for_each_populated_zone(zone) {
3771 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3775 for_each_online_cpu(cpu)
3776 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3784 " active_anon:%lukB"
3785 " inactive_anon:%lukB"
3786 " active_file:%lukB"
3787 " inactive_file:%lukB"
3788 " unevictable:%lukB"
3789 " isolated(anon):%lukB"
3790 " isolated(file):%lukB"
3798 " slab_reclaimable:%lukB"
3799 " slab_unreclaimable:%lukB"
3800 " kernel_stack:%lukB"
3807 " writeback_tmp:%lukB"
3808 " pages_scanned:%lu"
3809 " all_unreclaimable? %s"
3812 K(zone_page_state(zone, NR_FREE_PAGES)),
3813 K(min_wmark_pages(zone)),
3814 K(low_wmark_pages(zone)),
3815 K(high_wmark_pages(zone)),
3816 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3817 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3818 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3819 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3820 K(zone_page_state(zone, NR_UNEVICTABLE)),
3821 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3822 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3823 K(zone->present_pages),
3824 K(zone->managed_pages),
3825 K(zone_page_state(zone, NR_MLOCK)),
3826 K(zone_page_state(zone, NR_FILE_DIRTY)),
3827 K(zone_page_state(zone, NR_WRITEBACK)),
3828 K(zone_page_state(zone, NR_FILE_MAPPED)),
3829 K(zone_page_state(zone, NR_SHMEM)),
3830 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3831 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3832 zone_page_state(zone, NR_KERNEL_STACK) *
3834 K(zone_page_state(zone, NR_PAGETABLE)),
3835 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3836 K(zone_page_state(zone, NR_BOUNCE)),
3838 K(this_cpu_read(zone->pageset->pcp.count)),
3839 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3840 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3841 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3842 (!zone_reclaimable(zone) ? "yes" : "no")
3844 printk("lowmem_reserve[]:");
3845 for (i = 0; i < MAX_NR_ZONES; i++)
3846 printk(" %ld", zone->lowmem_reserve[i]);
3850 for_each_populated_zone(zone) {
3852 unsigned long nr[MAX_ORDER], flags, total = 0;
3853 unsigned char types[MAX_ORDER];
3855 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3858 printk("%s: ", zone->name);
3860 spin_lock_irqsave(&zone->lock, flags);
3861 for (order = 0; order < MAX_ORDER; order++) {
3862 struct free_area *area = &zone->free_area[order];
3865 nr[order] = area->nr_free;
3866 total += nr[order] << order;
3869 for (type = 0; type < MIGRATE_TYPES; type++) {
3870 if (!list_empty(&area->free_list[type]))
3871 types[order] |= 1 << type;
3874 spin_unlock_irqrestore(&zone->lock, flags);
3875 for (order = 0; order < MAX_ORDER; order++) {
3876 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3878 show_migration_types(types[order]);
3880 printk("= %lukB\n", K(total));
3883 hugetlb_show_meminfo();
3885 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3887 show_swap_cache_info();
3890 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3892 zoneref->zone = zone;
3893 zoneref->zone_idx = zone_idx(zone);
3897 * Builds allocation fallback zone lists.
3899 * Add all populated zones of a node to the zonelist.
3901 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3905 enum zone_type zone_type = MAX_NR_ZONES;
3909 zone = pgdat->node_zones + zone_type;
3910 if (populated_zone(zone)) {
3911 zoneref_set_zone(zone,
3912 &zonelist->_zonerefs[nr_zones++]);
3913 check_highest_zone(zone_type);
3915 } while (zone_type);
3923 * 0 = automatic detection of better ordering.
3924 * 1 = order by ([node] distance, -zonetype)
3925 * 2 = order by (-zonetype, [node] distance)
3927 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3928 * the same zonelist. So only NUMA can configure this param.
3930 #define ZONELIST_ORDER_DEFAULT 0
3931 #define ZONELIST_ORDER_NODE 1
3932 #define ZONELIST_ORDER_ZONE 2
3934 /* zonelist order in the kernel.
3935 * set_zonelist_order() will set this to NODE or ZONE.
3937 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3938 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3942 /* The value user specified ....changed by config */
3943 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3944 /* string for sysctl */
3945 #define NUMA_ZONELIST_ORDER_LEN 16
3946 char numa_zonelist_order[16] = "default";
3949 * interface for configure zonelist ordering.
3950 * command line option "numa_zonelist_order"
3951 * = "[dD]efault - default, automatic configuration.
3952 * = "[nN]ode - order by node locality, then by zone within node
3953 * = "[zZ]one - order by zone, then by locality within zone
3956 static int __parse_numa_zonelist_order(char *s)
3958 if (*s == 'd' || *s == 'D') {
3959 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3960 } else if (*s == 'n' || *s == 'N') {
3961 user_zonelist_order = ZONELIST_ORDER_NODE;
3962 } else if (*s == 'z' || *s == 'Z') {
3963 user_zonelist_order = ZONELIST_ORDER_ZONE;
3966 "Ignoring invalid numa_zonelist_order value: "
3973 static __init int setup_numa_zonelist_order(char *s)
3980 ret = __parse_numa_zonelist_order(s);
3982 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3986 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3989 * sysctl handler for numa_zonelist_order
3991 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3992 void __user *buffer, size_t *length,
3995 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3997 static DEFINE_MUTEX(zl_order_mutex);
3999 mutex_lock(&zl_order_mutex);
4001 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4005 strcpy(saved_string, (char *)table->data);
4007 ret = proc_dostring(table, write, buffer, length, ppos);
4011 int oldval = user_zonelist_order;
4013 ret = __parse_numa_zonelist_order((char *)table->data);
4016 * bogus value. restore saved string
4018 strncpy((char *)table->data, saved_string,
4019 NUMA_ZONELIST_ORDER_LEN);
4020 user_zonelist_order = oldval;
4021 } else if (oldval != user_zonelist_order) {
4022 mutex_lock(&zonelists_mutex);
4023 build_all_zonelists(NULL, NULL);
4024 mutex_unlock(&zonelists_mutex);
4028 mutex_unlock(&zl_order_mutex);
4033 #define MAX_NODE_LOAD (nr_online_nodes)
4034 static int node_load[MAX_NUMNODES];
4037 * find_next_best_node - find the next node that should appear in a given node's fallback list
4038 * @node: node whose fallback list we're appending
4039 * @used_node_mask: nodemask_t of already used nodes
4041 * We use a number of factors to determine which is the next node that should
4042 * appear on a given node's fallback list. The node should not have appeared
4043 * already in @node's fallback list, and it should be the next closest node
4044 * according to the distance array (which contains arbitrary distance values
4045 * from each node to each node in the system), and should also prefer nodes
4046 * with no CPUs, since presumably they'll have very little allocation pressure
4047 * on them otherwise.
4048 * It returns -1 if no node is found.
4050 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4053 int min_val = INT_MAX;
4054 int best_node = NUMA_NO_NODE;
4055 const struct cpumask *tmp = cpumask_of_node(0);
4057 /* Use the local node if we haven't already */
4058 if (!node_isset(node, *used_node_mask)) {
4059 node_set(node, *used_node_mask);
4063 for_each_node_state(n, N_MEMORY) {
4065 /* Don't want a node to appear more than once */
4066 if (node_isset(n, *used_node_mask))
4069 /* Use the distance array to find the distance */
4070 val = node_distance(node, n);
4072 /* Penalize nodes under us ("prefer the next node") */
4075 /* Give preference to headless and unused nodes */
4076 tmp = cpumask_of_node(n);
4077 if (!cpumask_empty(tmp))
4078 val += PENALTY_FOR_NODE_WITH_CPUS;
4080 /* Slight preference for less loaded node */
4081 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4082 val += node_load[n];
4084 if (val < min_val) {
4091 node_set(best_node, *used_node_mask);
4098 * Build zonelists ordered by node and zones within node.
4099 * This results in maximum locality--normal zone overflows into local
4100 * DMA zone, if any--but risks exhausting DMA zone.
4102 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4105 struct zonelist *zonelist;
4107 zonelist = &pgdat->node_zonelists[0];
4108 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4110 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4111 zonelist->_zonerefs[j].zone = NULL;
4112 zonelist->_zonerefs[j].zone_idx = 0;
4116 * Build gfp_thisnode zonelists
4118 static void build_thisnode_zonelists(pg_data_t *pgdat)
4121 struct zonelist *zonelist;
4123 zonelist = &pgdat->node_zonelists[1];
4124 j = build_zonelists_node(pgdat, zonelist, 0);
4125 zonelist->_zonerefs[j].zone = NULL;
4126 zonelist->_zonerefs[j].zone_idx = 0;
4130 * Build zonelists ordered by zone and nodes within zones.
4131 * This results in conserving DMA zone[s] until all Normal memory is
4132 * exhausted, but results in overflowing to remote node while memory
4133 * may still exist in local DMA zone.
4135 static int node_order[MAX_NUMNODES];
4137 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4140 int zone_type; /* needs to be signed */
4142 struct zonelist *zonelist;
4144 zonelist = &pgdat->node_zonelists[0];
4146 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4147 for (j = 0; j < nr_nodes; j++) {
4148 node = node_order[j];
4149 z = &NODE_DATA(node)->node_zones[zone_type];
4150 if (populated_zone(z)) {
4152 &zonelist->_zonerefs[pos++]);
4153 check_highest_zone(zone_type);
4157 zonelist->_zonerefs[pos].zone = NULL;
4158 zonelist->_zonerefs[pos].zone_idx = 0;
4161 #if defined(CONFIG_64BIT)
4163 * Devices that require DMA32/DMA are relatively rare and do not justify a
4164 * penalty to every machine in case the specialised case applies. Default
4165 * to Node-ordering on 64-bit NUMA machines
4167 static int default_zonelist_order(void)
4169 return ZONELIST_ORDER_NODE;
4173 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4174 * by the kernel. If processes running on node 0 deplete the low memory zone
4175 * then reclaim will occur more frequency increasing stalls and potentially
4176 * be easier to OOM if a large percentage of the zone is under writeback or
4177 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4178 * Hence, default to zone ordering on 32-bit.
4180 static int default_zonelist_order(void)
4182 return ZONELIST_ORDER_ZONE;
4184 #endif /* CONFIG_64BIT */
4186 static void set_zonelist_order(void)
4188 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4189 current_zonelist_order = default_zonelist_order();
4191 current_zonelist_order = user_zonelist_order;
4194 static void build_zonelists(pg_data_t *pgdat)
4198 nodemask_t used_mask;
4199 int local_node, prev_node;
4200 struct zonelist *zonelist;
4201 unsigned int order = current_zonelist_order;
4203 /* initialize zonelists */
4204 for (i = 0; i < MAX_ZONELISTS; i++) {
4205 zonelist = pgdat->node_zonelists + i;
4206 zonelist->_zonerefs[0].zone = NULL;
4207 zonelist->_zonerefs[0].zone_idx = 0;
4210 /* NUMA-aware ordering of nodes */
4211 local_node = pgdat->node_id;
4212 load = nr_online_nodes;
4213 prev_node = local_node;
4214 nodes_clear(used_mask);
4216 memset(node_order, 0, sizeof(node_order));
4219 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4221 * We don't want to pressure a particular node.
4222 * So adding penalty to the first node in same
4223 * distance group to make it round-robin.
4225 if (node_distance(local_node, node) !=
4226 node_distance(local_node, prev_node))
4227 node_load[node] = load;
4231 if (order == ZONELIST_ORDER_NODE)
4232 build_zonelists_in_node_order(pgdat, node);
4234 node_order[j++] = node; /* remember order */
4237 if (order == ZONELIST_ORDER_ZONE) {
4238 /* calculate node order -- i.e., DMA last! */
4239 build_zonelists_in_zone_order(pgdat, j);
4242 build_thisnode_zonelists(pgdat);
4245 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4247 * Return node id of node used for "local" allocations.
4248 * I.e., first node id of first zone in arg node's generic zonelist.
4249 * Used for initializing percpu 'numa_mem', which is used primarily
4250 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4252 int local_memory_node(int node)
4256 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4257 gfp_zone(GFP_KERNEL),
4264 #else /* CONFIG_NUMA */
4266 static void set_zonelist_order(void)
4268 current_zonelist_order = ZONELIST_ORDER_ZONE;
4271 static void build_zonelists(pg_data_t *pgdat)
4273 int node, local_node;
4275 struct zonelist *zonelist;
4277 local_node = pgdat->node_id;
4279 zonelist = &pgdat->node_zonelists[0];
4280 j = build_zonelists_node(pgdat, zonelist, 0);
4283 * Now we build the zonelist so that it contains the zones
4284 * of all the other nodes.
4285 * We don't want to pressure a particular node, so when
4286 * building the zones for node N, we make sure that the
4287 * zones coming right after the local ones are those from
4288 * node N+1 (modulo N)
4290 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4291 if (!node_online(node))
4293 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4295 for (node = 0; node < local_node; node++) {
4296 if (!node_online(node))
4298 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4301 zonelist->_zonerefs[j].zone = NULL;
4302 zonelist->_zonerefs[j].zone_idx = 0;
4305 #endif /* CONFIG_NUMA */
4308 * Boot pageset table. One per cpu which is going to be used for all
4309 * zones and all nodes. The parameters will be set in such a way
4310 * that an item put on a list will immediately be handed over to
4311 * the buddy list. This is safe since pageset manipulation is done
4312 * with interrupts disabled.
4314 * The boot_pagesets must be kept even after bootup is complete for
4315 * unused processors and/or zones. They do play a role for bootstrapping
4316 * hotplugged processors.
4318 * zoneinfo_show() and maybe other functions do
4319 * not check if the processor is online before following the pageset pointer.
4320 * Other parts of the kernel may not check if the zone is available.
4322 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4323 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4324 static void setup_zone_pageset(struct zone *zone);
4327 * Global mutex to protect against size modification of zonelists
4328 * as well as to serialize pageset setup for the new populated zone.
4330 DEFINE_MUTEX(zonelists_mutex);
4332 /* return values int ....just for stop_machine() */
4333 static int __build_all_zonelists(void *data)
4337 pg_data_t *self = data;
4340 memset(node_load, 0, sizeof(node_load));
4343 if (self && !node_online(self->node_id)) {
4344 build_zonelists(self);
4347 for_each_online_node(nid) {
4348 pg_data_t *pgdat = NODE_DATA(nid);
4350 build_zonelists(pgdat);
4354 * Initialize the boot_pagesets that are going to be used
4355 * for bootstrapping processors. The real pagesets for
4356 * each zone will be allocated later when the per cpu
4357 * allocator is available.
4359 * boot_pagesets are used also for bootstrapping offline
4360 * cpus if the system is already booted because the pagesets
4361 * are needed to initialize allocators on a specific cpu too.
4362 * F.e. the percpu allocator needs the page allocator which
4363 * needs the percpu allocator in order to allocate its pagesets
4364 * (a chicken-egg dilemma).
4366 for_each_possible_cpu(cpu) {
4367 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4369 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4371 * We now know the "local memory node" for each node--
4372 * i.e., the node of the first zone in the generic zonelist.
4373 * Set up numa_mem percpu variable for on-line cpus. During
4374 * boot, only the boot cpu should be on-line; we'll init the
4375 * secondary cpus' numa_mem as they come on-line. During
4376 * node/memory hotplug, we'll fixup all on-line cpus.
4378 if (cpu_online(cpu))
4379 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4386 static noinline void __init
4387 build_all_zonelists_init(void)
4389 __build_all_zonelists(NULL);
4390 mminit_verify_zonelist();
4391 cpuset_init_current_mems_allowed();
4395 * Called with zonelists_mutex held always
4396 * unless system_state == SYSTEM_BOOTING.
4398 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4399 * [we're only called with non-NULL zone through __meminit paths] and
4400 * (2) call of __init annotated helper build_all_zonelists_init
4401 * [protected by SYSTEM_BOOTING].
4403 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4405 set_zonelist_order();
4407 if (system_state == SYSTEM_BOOTING) {
4408 build_all_zonelists_init();
4410 #ifdef CONFIG_MEMORY_HOTPLUG
4412 setup_zone_pageset(zone);
4414 /* we have to stop all cpus to guarantee there is no user
4416 stop_machine(__build_all_zonelists, pgdat, NULL);
4417 /* cpuset refresh routine should be here */
4419 vm_total_pages = nr_free_pagecache_pages();
4421 * Disable grouping by mobility if the number of pages in the
4422 * system is too low to allow the mechanism to work. It would be
4423 * more accurate, but expensive to check per-zone. This check is
4424 * made on memory-hotadd so a system can start with mobility
4425 * disabled and enable it later
4427 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4428 page_group_by_mobility_disabled = 1;
4430 page_group_by_mobility_disabled = 0;
4432 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4433 "Total pages: %ld\n",
4435 zonelist_order_name[current_zonelist_order],
4436 page_group_by_mobility_disabled ? "off" : "on",
4439 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4444 * Helper functions to size the waitqueue hash table.
4445 * Essentially these want to choose hash table sizes sufficiently
4446 * large so that collisions trying to wait on pages are rare.
4447 * But in fact, the number of active page waitqueues on typical
4448 * systems is ridiculously low, less than 200. So this is even
4449 * conservative, even though it seems large.
4451 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4452 * waitqueues, i.e. the size of the waitq table given the number of pages.
4454 #define PAGES_PER_WAITQUEUE 256
4456 #ifndef CONFIG_MEMORY_HOTPLUG
4457 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4459 unsigned long size = 1;
4461 pages /= PAGES_PER_WAITQUEUE;
4463 while (size < pages)
4467 * Once we have dozens or even hundreds of threads sleeping
4468 * on IO we've got bigger problems than wait queue collision.
4469 * Limit the size of the wait table to a reasonable size.
4471 size = min(size, 4096UL);
4473 return max(size, 4UL);
4477 * A zone's size might be changed by hot-add, so it is not possible to determine
4478 * a suitable size for its wait_table. So we use the maximum size now.
4480 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4482 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4483 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4484 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4486 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4487 * or more by the traditional way. (See above). It equals:
4489 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4490 * ia64(16K page size) : = ( 8G + 4M)byte.
4491 * powerpc (64K page size) : = (32G +16M)byte.
4493 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4500 * This is an integer logarithm so that shifts can be used later
4501 * to extract the more random high bits from the multiplicative
4502 * hash function before the remainder is taken.
4504 static inline unsigned long wait_table_bits(unsigned long size)
4510 * Initially all pages are reserved - free ones are freed
4511 * up by free_all_bootmem() once the early boot process is
4512 * done. Non-atomic initialization, single-pass.
4514 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4515 unsigned long start_pfn, enum memmap_context context)
4517 pg_data_t *pgdat = NODE_DATA(nid);
4518 unsigned long end_pfn = start_pfn + size;
4521 unsigned long nr_initialised = 0;
4523 if (highest_memmap_pfn < end_pfn - 1)
4524 highest_memmap_pfn = end_pfn - 1;
4526 z = &pgdat->node_zones[zone];
4527 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4529 * There can be holes in boot-time mem_map[]s
4530 * handed to this function. They do not
4531 * exist on hotplugged memory.
4533 if (context == MEMMAP_EARLY) {
4534 if (!early_pfn_valid(pfn))
4536 if (!early_pfn_in_nid(pfn, nid))
4538 if (!update_defer_init(pgdat, pfn, end_pfn,
4544 * Mark the block movable so that blocks are reserved for
4545 * movable at startup. This will force kernel allocations
4546 * to reserve their blocks rather than leaking throughout
4547 * the address space during boot when many long-lived
4548 * kernel allocations are made.
4550 * bitmap is created for zone's valid pfn range. but memmap
4551 * can be created for invalid pages (for alignment)
4552 * check here not to call set_pageblock_migratetype() against
4555 if (!(pfn & (pageblock_nr_pages - 1))) {
4556 struct page *page = pfn_to_page(pfn);
4558 __init_single_page(page, pfn, zone, nid);
4559 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4561 __init_single_pfn(pfn, zone, nid);
4566 static void __meminit zone_init_free_lists(struct zone *zone)
4568 unsigned int order, t;
4569 for_each_migratetype_order(order, t) {
4570 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4571 zone->free_area[order].nr_free = 0;
4575 #ifndef __HAVE_ARCH_MEMMAP_INIT
4576 #define memmap_init(size, nid, zone, start_pfn) \
4577 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4580 static int zone_batchsize(struct zone *zone)
4586 * The per-cpu-pages pools are set to around 1000th of the
4587 * size of the zone. But no more than 1/2 of a meg.
4589 * OK, so we don't know how big the cache is. So guess.
4591 batch = zone->managed_pages / 1024;
4592 if (batch * PAGE_SIZE > 512 * 1024)
4593 batch = (512 * 1024) / PAGE_SIZE;
4594 batch /= 4; /* We effectively *= 4 below */
4599 * Clamp the batch to a 2^n - 1 value. Having a power
4600 * of 2 value was found to be more likely to have
4601 * suboptimal cache aliasing properties in some cases.
4603 * For example if 2 tasks are alternately allocating
4604 * batches of pages, one task can end up with a lot
4605 * of pages of one half of the possible page colors
4606 * and the other with pages of the other colors.
4608 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4613 /* The deferral and batching of frees should be suppressed under NOMMU
4616 * The problem is that NOMMU needs to be able to allocate large chunks
4617 * of contiguous memory as there's no hardware page translation to
4618 * assemble apparent contiguous memory from discontiguous pages.
4620 * Queueing large contiguous runs of pages for batching, however,
4621 * causes the pages to actually be freed in smaller chunks. As there
4622 * can be a significant delay between the individual batches being
4623 * recycled, this leads to the once large chunks of space being
4624 * fragmented and becoming unavailable for high-order allocations.
4631 * pcp->high and pcp->batch values are related and dependent on one another:
4632 * ->batch must never be higher then ->high.
4633 * The following function updates them in a safe manner without read side
4636 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4637 * those fields changing asynchronously (acording the the above rule).
4639 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4640 * outside of boot time (or some other assurance that no concurrent updaters
4643 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4644 unsigned long batch)
4646 /* start with a fail safe value for batch */
4650 /* Update high, then batch, in order */
4657 /* a companion to pageset_set_high() */
4658 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4660 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4663 static void pageset_init(struct per_cpu_pageset *p)
4665 struct per_cpu_pages *pcp;
4668 memset(p, 0, sizeof(*p));
4672 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4673 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4676 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4679 pageset_set_batch(p, batch);
4683 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4684 * to the value high for the pageset p.
4686 static void pageset_set_high(struct per_cpu_pageset *p,
4689 unsigned long batch = max(1UL, high / 4);
4690 if ((high / 4) > (PAGE_SHIFT * 8))
4691 batch = PAGE_SHIFT * 8;
4693 pageset_update(&p->pcp, high, batch);
4696 static void pageset_set_high_and_batch(struct zone *zone,
4697 struct per_cpu_pageset *pcp)
4699 if (percpu_pagelist_fraction)
4700 pageset_set_high(pcp,
4701 (zone->managed_pages /
4702 percpu_pagelist_fraction));
4704 pageset_set_batch(pcp, zone_batchsize(zone));
4707 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4709 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4712 pageset_set_high_and_batch(zone, pcp);
4715 static void __meminit setup_zone_pageset(struct zone *zone)
4718 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4719 for_each_possible_cpu(cpu)
4720 zone_pageset_init(zone, cpu);
4724 * Allocate per cpu pagesets and initialize them.
4725 * Before this call only boot pagesets were available.
4727 void __init setup_per_cpu_pageset(void)
4731 for_each_populated_zone(zone)
4732 setup_zone_pageset(zone);
4735 static noinline __init_refok
4736 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4742 * The per-page waitqueue mechanism uses hashed waitqueues
4745 zone->wait_table_hash_nr_entries =
4746 wait_table_hash_nr_entries(zone_size_pages);
4747 zone->wait_table_bits =
4748 wait_table_bits(zone->wait_table_hash_nr_entries);
4749 alloc_size = zone->wait_table_hash_nr_entries
4750 * sizeof(wait_queue_head_t);
4752 if (!slab_is_available()) {
4753 zone->wait_table = (wait_queue_head_t *)
4754 memblock_virt_alloc_node_nopanic(
4755 alloc_size, zone->zone_pgdat->node_id);
4758 * This case means that a zone whose size was 0 gets new memory
4759 * via memory hot-add.
4760 * But it may be the case that a new node was hot-added. In
4761 * this case vmalloc() will not be able to use this new node's
4762 * memory - this wait_table must be initialized to use this new
4763 * node itself as well.
4764 * To use this new node's memory, further consideration will be
4767 zone->wait_table = vmalloc(alloc_size);
4769 if (!zone->wait_table)
4772 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4773 init_waitqueue_head(zone->wait_table + i);
4778 static __meminit void zone_pcp_init(struct zone *zone)
4781 * per cpu subsystem is not up at this point. The following code
4782 * relies on the ability of the linker to provide the
4783 * offset of a (static) per cpu variable into the per cpu area.
4785 zone->pageset = &boot_pageset;
4787 if (populated_zone(zone))
4788 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4789 zone->name, zone->present_pages,
4790 zone_batchsize(zone));
4793 int __meminit init_currently_empty_zone(struct zone *zone,
4794 unsigned long zone_start_pfn,
4797 struct pglist_data *pgdat = zone->zone_pgdat;
4799 ret = zone_wait_table_init(zone, size);
4802 pgdat->nr_zones = zone_idx(zone) + 1;
4804 zone->zone_start_pfn = zone_start_pfn;
4806 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4807 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4809 (unsigned long)zone_idx(zone),
4810 zone_start_pfn, (zone_start_pfn + size));
4812 zone_init_free_lists(zone);
4817 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4818 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4821 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4823 int __meminit __early_pfn_to_nid(unsigned long pfn,
4824 struct mminit_pfnnid_cache *state)
4826 unsigned long start_pfn, end_pfn;
4829 if (state->last_start <= pfn && pfn < state->last_end)
4830 return state->last_nid;
4832 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4834 state->last_start = start_pfn;
4835 state->last_end = end_pfn;
4836 state->last_nid = nid;
4841 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4844 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4845 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4846 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4848 * If an architecture guarantees that all ranges registered contain no holes
4849 * and may be freed, this this function may be used instead of calling
4850 * memblock_free_early_nid() manually.
4852 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4854 unsigned long start_pfn, end_pfn;
4857 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4858 start_pfn = min(start_pfn, max_low_pfn);
4859 end_pfn = min(end_pfn, max_low_pfn);
4861 if (start_pfn < end_pfn)
4862 memblock_free_early_nid(PFN_PHYS(start_pfn),
4863 (end_pfn - start_pfn) << PAGE_SHIFT,
4869 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4870 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4872 * If an architecture guarantees that all ranges registered contain no holes and may
4873 * be freed, this function may be used instead of calling memory_present() manually.
4875 void __init sparse_memory_present_with_active_regions(int nid)
4877 unsigned long start_pfn, end_pfn;
4880 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4881 memory_present(this_nid, start_pfn, end_pfn);
4885 * get_pfn_range_for_nid - Return the start and end page frames for a node
4886 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4887 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4888 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4890 * It returns the start and end page frame of a node based on information
4891 * provided by memblock_set_node(). If called for a node
4892 * with no available memory, a warning is printed and the start and end
4895 void __meminit get_pfn_range_for_nid(unsigned int nid,
4896 unsigned long *start_pfn, unsigned long *end_pfn)
4898 unsigned long this_start_pfn, this_end_pfn;
4904 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4905 *start_pfn = min(*start_pfn, this_start_pfn);
4906 *end_pfn = max(*end_pfn, this_end_pfn);
4909 if (*start_pfn == -1UL)
4914 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4915 * assumption is made that zones within a node are ordered in monotonic
4916 * increasing memory addresses so that the "highest" populated zone is used
4918 static void __init find_usable_zone_for_movable(void)
4921 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4922 if (zone_index == ZONE_MOVABLE)
4925 if (arch_zone_highest_possible_pfn[zone_index] >
4926 arch_zone_lowest_possible_pfn[zone_index])
4930 VM_BUG_ON(zone_index == -1);
4931 movable_zone = zone_index;
4935 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4936 * because it is sized independent of architecture. Unlike the other zones,
4937 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4938 * in each node depending on the size of each node and how evenly kernelcore
4939 * is distributed. This helper function adjusts the zone ranges
4940 * provided by the architecture for a given node by using the end of the
4941 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4942 * zones within a node are in order of monotonic increases memory addresses
4944 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4945 unsigned long zone_type,
4946 unsigned long node_start_pfn,
4947 unsigned long node_end_pfn,
4948 unsigned long *zone_start_pfn,
4949 unsigned long *zone_end_pfn)
4951 /* Only adjust if ZONE_MOVABLE is on this node */
4952 if (zone_movable_pfn[nid]) {
4953 /* Size ZONE_MOVABLE */
4954 if (zone_type == ZONE_MOVABLE) {
4955 *zone_start_pfn = zone_movable_pfn[nid];
4956 *zone_end_pfn = min(node_end_pfn,
4957 arch_zone_highest_possible_pfn[movable_zone]);
4959 /* Adjust for ZONE_MOVABLE starting within this range */
4960 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4961 *zone_end_pfn > zone_movable_pfn[nid]) {
4962 *zone_end_pfn = zone_movable_pfn[nid];
4964 /* Check if this whole range is within ZONE_MOVABLE */
4965 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4966 *zone_start_pfn = *zone_end_pfn;
4971 * Return the number of pages a zone spans in a node, including holes
4972 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4974 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4975 unsigned long zone_type,
4976 unsigned long node_start_pfn,
4977 unsigned long node_end_pfn,
4978 unsigned long *ignored)
4980 unsigned long zone_start_pfn, zone_end_pfn;
4982 /* When hotadd a new node from cpu_up(), the node should be empty */
4983 if (!node_start_pfn && !node_end_pfn)
4986 /* Get the start and end of the zone */
4987 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4988 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4989 adjust_zone_range_for_zone_movable(nid, zone_type,
4990 node_start_pfn, node_end_pfn,
4991 &zone_start_pfn, &zone_end_pfn);
4993 /* Check that this node has pages within the zone's required range */
4994 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4997 /* Move the zone boundaries inside the node if necessary */
4998 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4999 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
5001 /* Return the spanned pages */
5002 return zone_end_pfn - zone_start_pfn;
5006 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5007 * then all holes in the requested range will be accounted for.
5009 unsigned long __meminit __absent_pages_in_range(int nid,
5010 unsigned long range_start_pfn,
5011 unsigned long range_end_pfn)
5013 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5014 unsigned long start_pfn, end_pfn;
5017 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5018 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5019 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5020 nr_absent -= end_pfn - start_pfn;
5026 * absent_pages_in_range - Return number of page frames in holes within a range
5027 * @start_pfn: The start PFN to start searching for holes
5028 * @end_pfn: The end PFN to stop searching for holes
5030 * It returns the number of pages frames in memory holes within a range.
5032 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5033 unsigned long end_pfn)
5035 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5038 /* Return the number of page frames in holes in a zone on a node */
5039 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5040 unsigned long zone_type,
5041 unsigned long node_start_pfn,
5042 unsigned long node_end_pfn,
5043 unsigned long *ignored)
5045 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5046 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5047 unsigned long zone_start_pfn, zone_end_pfn;
5049 /* When hotadd a new node from cpu_up(), the node should be empty */
5050 if (!node_start_pfn && !node_end_pfn)
5053 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5054 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5056 adjust_zone_range_for_zone_movable(nid, zone_type,
5057 node_start_pfn, node_end_pfn,
5058 &zone_start_pfn, &zone_end_pfn);
5059 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5062 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5063 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5064 unsigned long zone_type,
5065 unsigned long node_start_pfn,
5066 unsigned long node_end_pfn,
5067 unsigned long *zones_size)
5069 return zones_size[zone_type];
5072 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5073 unsigned long zone_type,
5074 unsigned long node_start_pfn,
5075 unsigned long node_end_pfn,
5076 unsigned long *zholes_size)
5081 return zholes_size[zone_type];
5084 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5086 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5087 unsigned long node_start_pfn,
5088 unsigned long node_end_pfn,
5089 unsigned long *zones_size,
5090 unsigned long *zholes_size)
5092 unsigned long realtotalpages = 0, totalpages = 0;
5095 for (i = 0; i < MAX_NR_ZONES; i++) {
5096 struct zone *zone = pgdat->node_zones + i;
5097 unsigned long size, real_size;
5099 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5103 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5104 node_start_pfn, node_end_pfn,
5106 zone->spanned_pages = size;
5107 zone->present_pages = real_size;
5110 realtotalpages += real_size;
5113 pgdat->node_spanned_pages = totalpages;
5114 pgdat->node_present_pages = realtotalpages;
5115 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5119 #ifndef CONFIG_SPARSEMEM
5121 * Calculate the size of the zone->blockflags rounded to an unsigned long
5122 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5123 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5124 * round what is now in bits to nearest long in bits, then return it in
5127 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5129 unsigned long usemapsize;
5131 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5132 usemapsize = roundup(zonesize, pageblock_nr_pages);
5133 usemapsize = usemapsize >> pageblock_order;
5134 usemapsize *= NR_PAGEBLOCK_BITS;
5135 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5137 return usemapsize / 8;
5140 static void __init setup_usemap(struct pglist_data *pgdat,
5142 unsigned long zone_start_pfn,
5143 unsigned long zonesize)
5145 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5146 zone->pageblock_flags = NULL;
5148 zone->pageblock_flags =
5149 memblock_virt_alloc_node_nopanic(usemapsize,
5153 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5154 unsigned long zone_start_pfn, unsigned long zonesize) {}
5155 #endif /* CONFIG_SPARSEMEM */
5157 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5159 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5160 void __paginginit set_pageblock_order(void)
5164 /* Check that pageblock_nr_pages has not already been setup */
5165 if (pageblock_order)
5168 if (HPAGE_SHIFT > PAGE_SHIFT)
5169 order = HUGETLB_PAGE_ORDER;
5171 order = MAX_ORDER - 1;
5174 * Assume the largest contiguous order of interest is a huge page.
5175 * This value may be variable depending on boot parameters on IA64 and
5178 pageblock_order = order;
5180 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5183 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5184 * is unused as pageblock_order is set at compile-time. See
5185 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5188 void __paginginit set_pageblock_order(void)
5192 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5194 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5195 unsigned long present_pages)
5197 unsigned long pages = spanned_pages;
5200 * Provide a more accurate estimation if there are holes within
5201 * the zone and SPARSEMEM is in use. If there are holes within the
5202 * zone, each populated memory region may cost us one or two extra
5203 * memmap pages due to alignment because memmap pages for each
5204 * populated regions may not naturally algined on page boundary.
5205 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5207 if (spanned_pages > present_pages + (present_pages >> 4) &&
5208 IS_ENABLED(CONFIG_SPARSEMEM))
5209 pages = present_pages;
5211 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5215 * Set up the zone data structures:
5216 * - mark all pages reserved
5217 * - mark all memory queues empty
5218 * - clear the memory bitmaps
5220 * NOTE: pgdat should get zeroed by caller.
5222 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5225 int nid = pgdat->node_id;
5226 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5229 pgdat_resize_init(pgdat);
5230 #ifdef CONFIG_NUMA_BALANCING
5231 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5232 pgdat->numabalancing_migrate_nr_pages = 0;
5233 pgdat->numabalancing_migrate_next_window = jiffies;
5235 init_waitqueue_head(&pgdat->kswapd_wait);
5236 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5237 pgdat_page_ext_init(pgdat);
5239 for (j = 0; j < MAX_NR_ZONES; j++) {
5240 struct zone *zone = pgdat->node_zones + j;
5241 unsigned long size, realsize, freesize, memmap_pages;
5243 size = zone->spanned_pages;
5244 realsize = freesize = zone->present_pages;
5247 * Adjust freesize so that it accounts for how much memory
5248 * is used by this zone for memmap. This affects the watermark
5249 * and per-cpu initialisations
5251 memmap_pages = calc_memmap_size(size, realsize);
5252 if (!is_highmem_idx(j)) {
5253 if (freesize >= memmap_pages) {
5254 freesize -= memmap_pages;
5257 " %s zone: %lu pages used for memmap\n",
5258 zone_names[j], memmap_pages);
5261 " %s zone: %lu pages exceeds freesize %lu\n",
5262 zone_names[j], memmap_pages, freesize);
5265 /* Account for reserved pages */
5266 if (j == 0 && freesize > dma_reserve) {
5267 freesize -= dma_reserve;
5268 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5269 zone_names[0], dma_reserve);
5272 if (!is_highmem_idx(j))
5273 nr_kernel_pages += freesize;
5274 /* Charge for highmem memmap if there are enough kernel pages */
5275 else if (nr_kernel_pages > memmap_pages * 2)
5276 nr_kernel_pages -= memmap_pages;
5277 nr_all_pages += freesize;
5280 * Set an approximate value for lowmem here, it will be adjusted
5281 * when the bootmem allocator frees pages into the buddy system.
5282 * And all highmem pages will be managed by the buddy system.
5284 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5287 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5289 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5291 zone->name = zone_names[j];
5292 spin_lock_init(&zone->lock);
5293 spin_lock_init(&zone->lru_lock);
5294 zone_seqlock_init(zone);
5295 zone->zone_pgdat = pgdat;
5296 zone_pcp_init(zone);
5298 /* For bootup, initialized properly in watermark setup */
5299 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5301 lruvec_init(&zone->lruvec);
5305 set_pageblock_order();
5306 setup_usemap(pgdat, zone, zone_start_pfn, size);
5307 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5309 memmap_init(size, nid, j, zone_start_pfn);
5310 zone_start_pfn += size;
5314 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5316 unsigned long __maybe_unused start = 0;
5317 unsigned long __maybe_unused offset = 0;
5319 /* Skip empty nodes */
5320 if (!pgdat->node_spanned_pages)
5323 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5324 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5325 offset = pgdat->node_start_pfn - start;
5326 /* ia64 gets its own node_mem_map, before this, without bootmem */
5327 if (!pgdat->node_mem_map) {
5328 unsigned long size, end;
5332 * The zone's endpoints aren't required to be MAX_ORDER
5333 * aligned but the node_mem_map endpoints must be in order
5334 * for the buddy allocator to function correctly.
5336 end = pgdat_end_pfn(pgdat);
5337 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5338 size = (end - start) * sizeof(struct page);
5339 map = alloc_remap(pgdat->node_id, size);
5341 map = memblock_virt_alloc_node_nopanic(size,
5343 pgdat->node_mem_map = map + offset;
5345 #ifndef CONFIG_NEED_MULTIPLE_NODES
5347 * With no DISCONTIG, the global mem_map is just set as node 0's
5349 if (pgdat == NODE_DATA(0)) {
5350 mem_map = NODE_DATA(0)->node_mem_map;
5351 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5352 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5354 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5357 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5360 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5361 unsigned long node_start_pfn, unsigned long *zholes_size)
5363 pg_data_t *pgdat = NODE_DATA(nid);
5364 unsigned long start_pfn = 0;
5365 unsigned long end_pfn = 0;
5367 /* pg_data_t should be reset to zero when it's allocated */
5368 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5370 pgdat->node_id = nid;
5371 pgdat->node_start_pfn = node_start_pfn;
5372 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5373 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5374 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5375 (u64)start_pfn << PAGE_SHIFT,
5376 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5378 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5379 zones_size, zholes_size);
5381 alloc_node_mem_map(pgdat);
5382 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5383 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5384 nid, (unsigned long)pgdat,
5385 (unsigned long)pgdat->node_mem_map);
5388 reset_deferred_meminit(pgdat);
5389 free_area_init_core(pgdat);
5392 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5394 #if MAX_NUMNODES > 1
5396 * Figure out the number of possible node ids.
5398 void __init setup_nr_node_ids(void)
5400 unsigned int highest;
5402 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5403 nr_node_ids = highest + 1;
5408 * node_map_pfn_alignment - determine the maximum internode alignment
5410 * This function should be called after node map is populated and sorted.
5411 * It calculates the maximum power of two alignment which can distinguish
5414 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5415 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5416 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5417 * shifted, 1GiB is enough and this function will indicate so.
5419 * This is used to test whether pfn -> nid mapping of the chosen memory
5420 * model has fine enough granularity to avoid incorrect mapping for the
5421 * populated node map.
5423 * Returns the determined alignment in pfn's. 0 if there is no alignment
5424 * requirement (single node).
5426 unsigned long __init node_map_pfn_alignment(void)
5428 unsigned long accl_mask = 0, last_end = 0;
5429 unsigned long start, end, mask;
5433 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5434 if (!start || last_nid < 0 || last_nid == nid) {
5441 * Start with a mask granular enough to pin-point to the
5442 * start pfn and tick off bits one-by-one until it becomes
5443 * too coarse to separate the current node from the last.
5445 mask = ~((1 << __ffs(start)) - 1);
5446 while (mask && last_end <= (start & (mask << 1)))
5449 /* accumulate all internode masks */
5453 /* convert mask to number of pages */
5454 return ~accl_mask + 1;
5457 /* Find the lowest pfn for a node */
5458 static unsigned long __init find_min_pfn_for_node(int nid)
5460 unsigned long min_pfn = ULONG_MAX;
5461 unsigned long start_pfn;
5464 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5465 min_pfn = min(min_pfn, start_pfn);
5467 if (min_pfn == ULONG_MAX) {
5469 "Could not find start_pfn for node %d\n", nid);
5477 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5479 * It returns the minimum PFN based on information provided via
5480 * memblock_set_node().
5482 unsigned long __init find_min_pfn_with_active_regions(void)
5484 return find_min_pfn_for_node(MAX_NUMNODES);
5488 * early_calculate_totalpages()
5489 * Sum pages in active regions for movable zone.
5490 * Populate N_MEMORY for calculating usable_nodes.
5492 static unsigned long __init early_calculate_totalpages(void)
5494 unsigned long totalpages = 0;
5495 unsigned long start_pfn, end_pfn;
5498 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5499 unsigned long pages = end_pfn - start_pfn;
5501 totalpages += pages;
5503 node_set_state(nid, N_MEMORY);
5509 * Find the PFN the Movable zone begins in each node. Kernel memory
5510 * is spread evenly between nodes as long as the nodes have enough
5511 * memory. When they don't, some nodes will have more kernelcore than
5514 static void __init find_zone_movable_pfns_for_nodes(void)
5517 unsigned long usable_startpfn;
5518 unsigned long kernelcore_node, kernelcore_remaining;
5519 /* save the state before borrow the nodemask */
5520 nodemask_t saved_node_state = node_states[N_MEMORY];
5521 unsigned long totalpages = early_calculate_totalpages();
5522 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5523 struct memblock_region *r;
5525 /* Need to find movable_zone earlier when movable_node is specified. */
5526 find_usable_zone_for_movable();
5529 * If movable_node is specified, ignore kernelcore and movablecore
5532 if (movable_node_is_enabled()) {
5533 for_each_memblock(memory, r) {
5534 if (!memblock_is_hotpluggable(r))
5539 usable_startpfn = PFN_DOWN(r->base);
5540 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5541 min(usable_startpfn, zone_movable_pfn[nid]) :
5549 * If movablecore=nn[KMG] was specified, calculate what size of
5550 * kernelcore that corresponds so that memory usable for
5551 * any allocation type is evenly spread. If both kernelcore
5552 * and movablecore are specified, then the value of kernelcore
5553 * will be used for required_kernelcore if it's greater than
5554 * what movablecore would have allowed.
5556 if (required_movablecore) {
5557 unsigned long corepages;
5560 * Round-up so that ZONE_MOVABLE is at least as large as what
5561 * was requested by the user
5563 required_movablecore =
5564 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5565 required_movablecore = min(totalpages, required_movablecore);
5566 corepages = totalpages - required_movablecore;
5568 required_kernelcore = max(required_kernelcore, corepages);
5572 * If kernelcore was not specified or kernelcore size is larger
5573 * than totalpages, there is no ZONE_MOVABLE.
5575 if (!required_kernelcore || required_kernelcore >= totalpages)
5578 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5579 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5582 /* Spread kernelcore memory as evenly as possible throughout nodes */
5583 kernelcore_node = required_kernelcore / usable_nodes;
5584 for_each_node_state(nid, N_MEMORY) {
5585 unsigned long start_pfn, end_pfn;
5588 * Recalculate kernelcore_node if the division per node
5589 * now exceeds what is necessary to satisfy the requested
5590 * amount of memory for the kernel
5592 if (required_kernelcore < kernelcore_node)
5593 kernelcore_node = required_kernelcore / usable_nodes;
5596 * As the map is walked, we track how much memory is usable
5597 * by the kernel using kernelcore_remaining. When it is
5598 * 0, the rest of the node is usable by ZONE_MOVABLE
5600 kernelcore_remaining = kernelcore_node;
5602 /* Go through each range of PFNs within this node */
5603 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5604 unsigned long size_pages;
5606 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5607 if (start_pfn >= end_pfn)
5610 /* Account for what is only usable for kernelcore */
5611 if (start_pfn < usable_startpfn) {
5612 unsigned long kernel_pages;
5613 kernel_pages = min(end_pfn, usable_startpfn)
5616 kernelcore_remaining -= min(kernel_pages,
5617 kernelcore_remaining);
5618 required_kernelcore -= min(kernel_pages,
5619 required_kernelcore);
5621 /* Continue if range is now fully accounted */
5622 if (end_pfn <= usable_startpfn) {
5625 * Push zone_movable_pfn to the end so
5626 * that if we have to rebalance
5627 * kernelcore across nodes, we will
5628 * not double account here
5630 zone_movable_pfn[nid] = end_pfn;
5633 start_pfn = usable_startpfn;
5637 * The usable PFN range for ZONE_MOVABLE is from
5638 * start_pfn->end_pfn. Calculate size_pages as the
5639 * number of pages used as kernelcore
5641 size_pages = end_pfn - start_pfn;
5642 if (size_pages > kernelcore_remaining)
5643 size_pages = kernelcore_remaining;
5644 zone_movable_pfn[nid] = start_pfn + size_pages;
5647 * Some kernelcore has been met, update counts and
5648 * break if the kernelcore for this node has been
5651 required_kernelcore -= min(required_kernelcore,
5653 kernelcore_remaining -= size_pages;
5654 if (!kernelcore_remaining)
5660 * If there is still required_kernelcore, we do another pass with one
5661 * less node in the count. This will push zone_movable_pfn[nid] further
5662 * along on the nodes that still have memory until kernelcore is
5666 if (usable_nodes && required_kernelcore > usable_nodes)
5670 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5671 for (nid = 0; nid < MAX_NUMNODES; nid++)
5672 zone_movable_pfn[nid] =
5673 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5676 /* restore the node_state */
5677 node_states[N_MEMORY] = saved_node_state;
5680 /* Any regular or high memory on that node ? */
5681 static void check_for_memory(pg_data_t *pgdat, int nid)
5683 enum zone_type zone_type;
5685 if (N_MEMORY == N_NORMAL_MEMORY)
5688 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5689 struct zone *zone = &pgdat->node_zones[zone_type];
5690 if (populated_zone(zone)) {
5691 node_set_state(nid, N_HIGH_MEMORY);
5692 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5693 zone_type <= ZONE_NORMAL)
5694 node_set_state(nid, N_NORMAL_MEMORY);
5701 * free_area_init_nodes - Initialise all pg_data_t and zone data
5702 * @max_zone_pfn: an array of max PFNs for each zone
5704 * This will call free_area_init_node() for each active node in the system.
5705 * Using the page ranges provided by memblock_set_node(), the size of each
5706 * zone in each node and their holes is calculated. If the maximum PFN
5707 * between two adjacent zones match, it is assumed that the zone is empty.
5708 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5709 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5710 * starts where the previous one ended. For example, ZONE_DMA32 starts
5711 * at arch_max_dma_pfn.
5713 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5715 unsigned long start_pfn, end_pfn;
5718 /* Record where the zone boundaries are */
5719 memset(arch_zone_lowest_possible_pfn, 0,
5720 sizeof(arch_zone_lowest_possible_pfn));
5721 memset(arch_zone_highest_possible_pfn, 0,
5722 sizeof(arch_zone_highest_possible_pfn));
5724 start_pfn = find_min_pfn_with_active_regions();
5726 for (i = 0; i < MAX_NR_ZONES; i++) {
5727 if (i == ZONE_MOVABLE)
5730 end_pfn = max(max_zone_pfn[i], start_pfn);
5731 arch_zone_lowest_possible_pfn[i] = start_pfn;
5732 arch_zone_highest_possible_pfn[i] = end_pfn;
5734 start_pfn = end_pfn;
5736 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5737 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5739 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5740 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5741 find_zone_movable_pfns_for_nodes();
5743 /* Print out the zone ranges */
5744 pr_info("Zone ranges:\n");
5745 for (i = 0; i < MAX_NR_ZONES; i++) {
5746 if (i == ZONE_MOVABLE)
5748 pr_info(" %-8s ", zone_names[i]);
5749 if (arch_zone_lowest_possible_pfn[i] ==
5750 arch_zone_highest_possible_pfn[i])
5753 pr_cont("[mem %#018Lx-%#018Lx]\n",
5754 (u64)arch_zone_lowest_possible_pfn[i]
5756 ((u64)arch_zone_highest_possible_pfn[i]
5757 << PAGE_SHIFT) - 1);
5760 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5761 pr_info("Movable zone start for each node\n");
5762 for (i = 0; i < MAX_NUMNODES; i++) {
5763 if (zone_movable_pfn[i])
5764 pr_info(" Node %d: %#018Lx\n", i,
5765 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5768 /* Print out the early node map */
5769 pr_info("Early memory node ranges\n");
5770 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5771 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5772 (u64)start_pfn << PAGE_SHIFT,
5773 ((u64)end_pfn << PAGE_SHIFT) - 1);
5775 /* Initialise every node */
5776 mminit_verify_pageflags_layout();
5777 setup_nr_node_ids();
5778 for_each_online_node(nid) {
5779 pg_data_t *pgdat = NODE_DATA(nid);
5780 free_area_init_node(nid, NULL,
5781 find_min_pfn_for_node(nid), NULL);
5783 /* Any memory on that node */
5784 if (pgdat->node_present_pages)
5785 node_set_state(nid, N_MEMORY);
5786 check_for_memory(pgdat, nid);
5790 static int __init cmdline_parse_core(char *p, unsigned long *core)
5792 unsigned long long coremem;
5796 coremem = memparse(p, &p);
5797 *core = coremem >> PAGE_SHIFT;
5799 /* Paranoid check that UL is enough for the coremem value */
5800 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5806 * kernelcore=size sets the amount of memory for use for allocations that
5807 * cannot be reclaimed or migrated.
5809 static int __init cmdline_parse_kernelcore(char *p)
5811 return cmdline_parse_core(p, &required_kernelcore);
5815 * movablecore=size sets the amount of memory for use for allocations that
5816 * can be reclaimed or migrated.
5818 static int __init cmdline_parse_movablecore(char *p)
5820 return cmdline_parse_core(p, &required_movablecore);
5823 early_param("kernelcore", cmdline_parse_kernelcore);
5824 early_param("movablecore", cmdline_parse_movablecore);
5826 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5828 void adjust_managed_page_count(struct page *page, long count)
5830 spin_lock(&managed_page_count_lock);
5831 page_zone(page)->managed_pages += count;
5832 totalram_pages += count;
5833 #ifdef CONFIG_HIGHMEM
5834 if (PageHighMem(page))
5835 totalhigh_pages += count;
5837 spin_unlock(&managed_page_count_lock);
5839 EXPORT_SYMBOL(adjust_managed_page_count);
5841 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5844 unsigned long pages = 0;
5846 start = (void *)PAGE_ALIGN((unsigned long)start);
5847 end = (void *)((unsigned long)end & PAGE_MASK);
5848 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5849 if ((unsigned int)poison <= 0xFF)
5850 memset(pos, poison, PAGE_SIZE);
5851 free_reserved_page(virt_to_page(pos));
5855 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5856 s, pages << (PAGE_SHIFT - 10), start, end);
5860 EXPORT_SYMBOL(free_reserved_area);
5862 #ifdef CONFIG_HIGHMEM
5863 void free_highmem_page(struct page *page)
5865 __free_reserved_page(page);
5867 page_zone(page)->managed_pages++;
5873 void __init mem_init_print_info(const char *str)
5875 unsigned long physpages, codesize, datasize, rosize, bss_size;
5876 unsigned long init_code_size, init_data_size;
5878 physpages = get_num_physpages();
5879 codesize = _etext - _stext;
5880 datasize = _edata - _sdata;
5881 rosize = __end_rodata - __start_rodata;
5882 bss_size = __bss_stop - __bss_start;
5883 init_data_size = __init_end - __init_begin;
5884 init_code_size = _einittext - _sinittext;
5887 * Detect special cases and adjust section sizes accordingly:
5888 * 1) .init.* may be embedded into .data sections
5889 * 2) .init.text.* may be out of [__init_begin, __init_end],
5890 * please refer to arch/tile/kernel/vmlinux.lds.S.
5891 * 3) .rodata.* may be embedded into .text or .data sections.
5893 #define adj_init_size(start, end, size, pos, adj) \
5895 if (start <= pos && pos < end && size > adj) \
5899 adj_init_size(__init_begin, __init_end, init_data_size,
5900 _sinittext, init_code_size);
5901 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5902 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5903 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5904 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5906 #undef adj_init_size
5908 pr_info("Memory: %luK/%luK available "
5909 "(%luK kernel code, %luK rwdata, %luK rodata, "
5910 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5911 #ifdef CONFIG_HIGHMEM
5915 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5916 codesize >> 10, datasize >> 10, rosize >> 10,
5917 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5918 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5919 totalcma_pages << (PAGE_SHIFT-10),
5920 #ifdef CONFIG_HIGHMEM
5921 totalhigh_pages << (PAGE_SHIFT-10),
5923 str ? ", " : "", str ? str : "");
5927 * set_dma_reserve - set the specified number of pages reserved in the first zone
5928 * @new_dma_reserve: The number of pages to mark reserved
5930 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5931 * In the DMA zone, a significant percentage may be consumed by kernel image
5932 * and other unfreeable allocations which can skew the watermarks badly. This
5933 * function may optionally be used to account for unfreeable pages in the
5934 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5935 * smaller per-cpu batchsize.
5937 void __init set_dma_reserve(unsigned long new_dma_reserve)
5939 dma_reserve = new_dma_reserve;
5942 void __init free_area_init(unsigned long *zones_size)
5944 free_area_init_node(0, zones_size,
5945 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5948 static int page_alloc_cpu_notify(struct notifier_block *self,
5949 unsigned long action, void *hcpu)
5951 int cpu = (unsigned long)hcpu;
5953 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5954 lru_add_drain_cpu(cpu);
5958 * Spill the event counters of the dead processor
5959 * into the current processors event counters.
5960 * This artificially elevates the count of the current
5963 vm_events_fold_cpu(cpu);
5966 * Zero the differential counters of the dead processor
5967 * so that the vm statistics are consistent.
5969 * This is only okay since the processor is dead and cannot
5970 * race with what we are doing.
5972 cpu_vm_stats_fold(cpu);
5977 void __init page_alloc_init(void)
5979 hotcpu_notifier(page_alloc_cpu_notify, 0);
5983 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5984 * or min_free_kbytes changes.
5986 static void calculate_totalreserve_pages(void)
5988 struct pglist_data *pgdat;
5989 unsigned long reserve_pages = 0;
5990 enum zone_type i, j;
5992 for_each_online_pgdat(pgdat) {
5993 for (i = 0; i < MAX_NR_ZONES; i++) {
5994 struct zone *zone = pgdat->node_zones + i;
5997 /* Find valid and maximum lowmem_reserve in the zone */
5998 for (j = i; j < MAX_NR_ZONES; j++) {
5999 if (zone->lowmem_reserve[j] > max)
6000 max = zone->lowmem_reserve[j];
6003 /* we treat the high watermark as reserved pages. */
6004 max += high_wmark_pages(zone);
6006 if (max > zone->managed_pages)
6007 max = zone->managed_pages;
6009 zone->totalreserve_pages = max;
6011 reserve_pages += max;
6014 totalreserve_pages = reserve_pages;
6018 * setup_per_zone_lowmem_reserve - called whenever
6019 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6020 * has a correct pages reserved value, so an adequate number of
6021 * pages are left in the zone after a successful __alloc_pages().
6023 static void setup_per_zone_lowmem_reserve(void)
6025 struct pglist_data *pgdat;
6026 enum zone_type j, idx;
6028 for_each_online_pgdat(pgdat) {
6029 for (j = 0; j < MAX_NR_ZONES; j++) {
6030 struct zone *zone = pgdat->node_zones + j;
6031 unsigned long managed_pages = zone->managed_pages;
6033 zone->lowmem_reserve[j] = 0;
6037 struct zone *lower_zone;
6041 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6042 sysctl_lowmem_reserve_ratio[idx] = 1;
6044 lower_zone = pgdat->node_zones + idx;
6045 lower_zone->lowmem_reserve[j] = managed_pages /
6046 sysctl_lowmem_reserve_ratio[idx];
6047 managed_pages += lower_zone->managed_pages;
6052 /* update totalreserve_pages */
6053 calculate_totalreserve_pages();
6056 static void __setup_per_zone_wmarks(void)
6058 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6059 unsigned long pages_low = extra_free_kbytes >> (PAGE_SHIFT - 10);
6060 unsigned long lowmem_pages = 0;
6062 unsigned long flags;
6064 /* Calculate total number of !ZONE_HIGHMEM pages */
6065 for_each_zone(zone) {
6066 if (!is_highmem(zone))
6067 lowmem_pages += zone->managed_pages;
6070 for_each_zone(zone) {
6073 spin_lock_irqsave(&zone->lock, flags);
6074 min = (u64)pages_min * zone->managed_pages;
6075 do_div(min, lowmem_pages);
6076 low = (u64)pages_low * zone->managed_pages;
6077 do_div(low, vm_total_pages);
6079 if (is_highmem(zone)) {
6081 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6082 * need highmem pages, so cap pages_min to a small
6085 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6086 * deltas control asynch page reclaim, and so should
6087 * not be capped for highmem.
6089 unsigned long min_pages;
6091 min_pages = zone->managed_pages / 1024;
6092 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6093 zone->watermark[WMARK_MIN] = min_pages;
6096 * If it's a lowmem zone, reserve a number of pages
6097 * proportionate to the zone's size.
6099 zone->watermark[WMARK_MIN] = min;
6102 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) +
6104 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) +
6107 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6108 high_wmark_pages(zone) - low_wmark_pages(zone) -
6109 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6111 spin_unlock_irqrestore(&zone->lock, flags);
6114 /* update totalreserve_pages */
6115 calculate_totalreserve_pages();
6119 * setup_per_zone_wmarks - called when min_free_kbytes changes
6120 * or when memory is hot-{added|removed}
6122 * Ensures that the watermark[min,low,high] values for each zone are set
6123 * correctly with respect to min_free_kbytes.
6125 void setup_per_zone_wmarks(void)
6127 mutex_lock(&zonelists_mutex);
6128 __setup_per_zone_wmarks();
6129 mutex_unlock(&zonelists_mutex);
6133 * The inactive anon list should be small enough that the VM never has to
6134 * do too much work, but large enough that each inactive page has a chance
6135 * to be referenced again before it is swapped out.
6137 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6138 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6139 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6140 * the anonymous pages are kept on the inactive list.
6143 * memory ratio inactive anon
6144 * -------------------------------------
6153 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6155 unsigned int gb, ratio;
6157 /* Zone size in gigabytes */
6158 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6160 ratio = int_sqrt(10 * gb);
6164 zone->inactive_ratio = ratio;
6167 static void __meminit setup_per_zone_inactive_ratio(void)
6172 calculate_zone_inactive_ratio(zone);
6176 * Initialise min_free_kbytes.
6178 * For small machines we want it small (128k min). For large machines
6179 * we want it large (64MB max). But it is not linear, because network
6180 * bandwidth does not increase linearly with machine size. We use
6182 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6183 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6199 int __meminit init_per_zone_wmark_min(void)
6201 unsigned long lowmem_kbytes;
6202 int new_min_free_kbytes;
6204 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6205 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6207 if (new_min_free_kbytes > user_min_free_kbytes) {
6208 min_free_kbytes = new_min_free_kbytes;
6209 if (min_free_kbytes < 128)
6210 min_free_kbytes = 128;
6211 if (min_free_kbytes > 65536)
6212 min_free_kbytes = 65536;
6214 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6215 new_min_free_kbytes, user_min_free_kbytes);
6217 setup_per_zone_wmarks();
6218 refresh_zone_stat_thresholds();
6219 setup_per_zone_lowmem_reserve();
6220 setup_per_zone_inactive_ratio();
6223 core_initcall(init_per_zone_wmark_min)
6226 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6227 * that we can call two helper functions whenever min_free_kbytes
6228 * or extra_free_kbytes changes.
6230 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6231 void __user *buffer, size_t *length, loff_t *ppos)
6235 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6240 user_min_free_kbytes = min_free_kbytes;
6241 setup_per_zone_wmarks();
6247 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6248 void __user *buffer, size_t *length, loff_t *ppos)
6253 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6258 zone->min_unmapped_pages = (zone->managed_pages *
6259 sysctl_min_unmapped_ratio) / 100;
6263 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6264 void __user *buffer, size_t *length, loff_t *ppos)
6269 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6274 zone->min_slab_pages = (zone->managed_pages *
6275 sysctl_min_slab_ratio) / 100;
6281 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6282 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6283 * whenever sysctl_lowmem_reserve_ratio changes.
6285 * The reserve ratio obviously has absolutely no relation with the
6286 * minimum watermarks. The lowmem reserve ratio can only make sense
6287 * if in function of the boot time zone sizes.
6289 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6290 void __user *buffer, size_t *length, loff_t *ppos)
6292 proc_dointvec_minmax(table, write, buffer, length, ppos);
6293 setup_per_zone_lowmem_reserve();
6298 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6299 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6300 * pagelist can have before it gets flushed back to buddy allocator.
6302 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6303 void __user *buffer, size_t *length, loff_t *ppos)
6306 int old_percpu_pagelist_fraction;
6309 mutex_lock(&pcp_batch_high_lock);
6310 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6312 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6313 if (!write || ret < 0)
6316 /* Sanity checking to avoid pcp imbalance */
6317 if (percpu_pagelist_fraction &&
6318 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6319 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6325 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6328 for_each_populated_zone(zone) {
6331 for_each_possible_cpu(cpu)
6332 pageset_set_high_and_batch(zone,
6333 per_cpu_ptr(zone->pageset, cpu));
6336 mutex_unlock(&pcp_batch_high_lock);
6341 int hashdist = HASHDIST_DEFAULT;
6343 static int __init set_hashdist(char *str)
6347 hashdist = simple_strtoul(str, &str, 0);
6350 __setup("hashdist=", set_hashdist);
6354 * allocate a large system hash table from bootmem
6355 * - it is assumed that the hash table must contain an exact power-of-2
6356 * quantity of entries
6357 * - limit is the number of hash buckets, not the total allocation size
6359 void *__init alloc_large_system_hash(const char *tablename,
6360 unsigned long bucketsize,
6361 unsigned long numentries,
6364 unsigned int *_hash_shift,
6365 unsigned int *_hash_mask,
6366 unsigned long low_limit,
6367 unsigned long high_limit)
6369 unsigned long long max = high_limit;
6370 unsigned long log2qty, size;
6373 /* allow the kernel cmdline to have a say */
6375 /* round applicable memory size up to nearest megabyte */
6376 numentries = nr_kernel_pages;
6378 /* It isn't necessary when PAGE_SIZE >= 1MB */
6379 if (PAGE_SHIFT < 20)
6380 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6382 /* limit to 1 bucket per 2^scale bytes of low memory */
6383 if (scale > PAGE_SHIFT)
6384 numentries >>= (scale - PAGE_SHIFT);
6386 numentries <<= (PAGE_SHIFT - scale);
6388 /* Make sure we've got at least a 0-order allocation.. */
6389 if (unlikely(flags & HASH_SMALL)) {
6390 /* Makes no sense without HASH_EARLY */
6391 WARN_ON(!(flags & HASH_EARLY));
6392 if (!(numentries >> *_hash_shift)) {
6393 numentries = 1UL << *_hash_shift;
6394 BUG_ON(!numentries);
6396 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6397 numentries = PAGE_SIZE / bucketsize;
6399 numentries = roundup_pow_of_two(numentries);
6401 /* limit allocation size to 1/16 total memory by default */
6403 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6404 do_div(max, bucketsize);
6406 max = min(max, 0x80000000ULL);
6408 if (numentries < low_limit)
6409 numentries = low_limit;
6410 if (numentries > max)
6413 log2qty = ilog2(numentries);
6416 size = bucketsize << log2qty;
6417 if (flags & HASH_EARLY)
6418 table = memblock_virt_alloc_nopanic(size, 0);
6420 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6423 * If bucketsize is not a power-of-two, we may free
6424 * some pages at the end of hash table which
6425 * alloc_pages_exact() automatically does
6427 if (get_order(size) < MAX_ORDER) {
6428 table = alloc_pages_exact(size, GFP_ATOMIC);
6429 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6432 } while (!table && size > PAGE_SIZE && --log2qty);
6435 panic("Failed to allocate %s hash table\n", tablename);
6437 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6440 ilog2(size) - PAGE_SHIFT,
6444 *_hash_shift = log2qty;
6446 *_hash_mask = (1 << log2qty) - 1;
6451 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6452 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6455 #ifdef CONFIG_SPARSEMEM
6456 return __pfn_to_section(pfn)->pageblock_flags;
6458 return zone->pageblock_flags;
6459 #endif /* CONFIG_SPARSEMEM */
6462 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6464 #ifdef CONFIG_SPARSEMEM
6465 pfn &= (PAGES_PER_SECTION-1);
6466 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6468 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6469 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6470 #endif /* CONFIG_SPARSEMEM */
6474 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6475 * @page: The page within the block of interest
6476 * @pfn: The target page frame number
6477 * @end_bitidx: The last bit of interest to retrieve
6478 * @mask: mask of bits that the caller is interested in
6480 * Return: pageblock_bits flags
6482 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6483 unsigned long end_bitidx,
6487 unsigned long *bitmap;
6488 unsigned long bitidx, word_bitidx;
6491 zone = page_zone(page);
6492 bitmap = get_pageblock_bitmap(zone, pfn);
6493 bitidx = pfn_to_bitidx(zone, pfn);
6494 word_bitidx = bitidx / BITS_PER_LONG;
6495 bitidx &= (BITS_PER_LONG-1);
6497 word = bitmap[word_bitidx];
6498 bitidx += end_bitidx;
6499 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6503 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6504 * @page: The page within the block of interest
6505 * @flags: The flags to set
6506 * @pfn: The target page frame number
6507 * @end_bitidx: The last bit of interest
6508 * @mask: mask of bits that the caller is interested in
6510 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6512 unsigned long end_bitidx,
6516 unsigned long *bitmap;
6517 unsigned long bitidx, word_bitidx;
6518 unsigned long old_word, word;
6520 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6522 zone = page_zone(page);
6523 bitmap = get_pageblock_bitmap(zone, pfn);
6524 bitidx = pfn_to_bitidx(zone, pfn);
6525 word_bitidx = bitidx / BITS_PER_LONG;
6526 bitidx &= (BITS_PER_LONG-1);
6528 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6530 bitidx += end_bitidx;
6531 mask <<= (BITS_PER_LONG - bitidx - 1);
6532 flags <<= (BITS_PER_LONG - bitidx - 1);
6534 word = READ_ONCE(bitmap[word_bitidx]);
6536 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6537 if (word == old_word)
6544 * This function checks whether pageblock includes unmovable pages or not.
6545 * If @count is not zero, it is okay to include less @count unmovable pages
6547 * PageLRU check without isolation or lru_lock could race so that
6548 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6549 * expect this function should be exact.
6551 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6552 bool skip_hwpoisoned_pages)
6554 unsigned long pfn, iter, found;
6558 * For avoiding noise data, lru_add_drain_all() should be called
6559 * If ZONE_MOVABLE, the zone never contains unmovable pages
6561 if (zone_idx(zone) == ZONE_MOVABLE)
6563 mt = get_pageblock_migratetype(page);
6564 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6567 pfn = page_to_pfn(page);
6568 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6569 unsigned long check = pfn + iter;
6571 if (!pfn_valid_within(check))
6574 page = pfn_to_page(check);
6577 * Hugepages are not in LRU lists, but they're movable.
6578 * We need not scan over tail pages bacause we don't
6579 * handle each tail page individually in migration.
6581 if (PageHuge(page)) {
6582 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6587 * We can't use page_count without pin a page
6588 * because another CPU can free compound page.
6589 * This check already skips compound tails of THP
6590 * because their page->_count is zero at all time.
6592 if (!atomic_read(&page->_count)) {
6593 if (PageBuddy(page))
6594 iter += (1 << page_order(page)) - 1;
6599 * The HWPoisoned page may be not in buddy system, and
6600 * page_count() is not 0.
6602 if (skip_hwpoisoned_pages && PageHWPoison(page))
6608 * If there are RECLAIMABLE pages, we need to check
6609 * it. But now, memory offline itself doesn't call
6610 * shrink_node_slabs() and it still to be fixed.
6613 * If the page is not RAM, page_count()should be 0.
6614 * we don't need more check. This is an _used_ not-movable page.
6616 * The problematic thing here is PG_reserved pages. PG_reserved
6617 * is set to both of a memory hole page and a _used_ kernel
6626 bool is_pageblock_removable_nolock(struct page *page)
6632 * We have to be careful here because we are iterating over memory
6633 * sections which are not zone aware so we might end up outside of
6634 * the zone but still within the section.
6635 * We have to take care about the node as well. If the node is offline
6636 * its NODE_DATA will be NULL - see page_zone.
6638 if (!node_online(page_to_nid(page)))
6641 zone = page_zone(page);
6642 pfn = page_to_pfn(page);
6643 if (!zone_spans_pfn(zone, pfn))
6646 return !has_unmovable_pages(zone, page, 0, true);
6651 static unsigned long pfn_max_align_down(unsigned long pfn)
6653 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6654 pageblock_nr_pages) - 1);
6657 static unsigned long pfn_max_align_up(unsigned long pfn)
6659 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6660 pageblock_nr_pages));
6663 /* [start, end) must belong to a single zone. */
6664 static int __alloc_contig_migrate_range(struct compact_control *cc,
6665 unsigned long start, unsigned long end)
6667 /* This function is based on compact_zone() from compaction.c. */
6668 unsigned long nr_reclaimed;
6669 unsigned long pfn = start;
6670 unsigned int tries = 0;
6675 while (pfn < end || !list_empty(&cc->migratepages)) {
6676 if (fatal_signal_pending(current)) {
6681 if (list_empty(&cc->migratepages)) {
6682 cc->nr_migratepages = 0;
6683 pfn = isolate_migratepages_range(cc, pfn, end);
6689 } else if (++tries == 5) {
6690 ret = ret < 0 ? ret : -EBUSY;
6694 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6696 cc->nr_migratepages -= nr_reclaimed;
6698 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6699 NULL, 0, cc->mode, MR_CMA);
6702 putback_movable_pages(&cc->migratepages);
6709 * alloc_contig_range() -- tries to allocate given range of pages
6710 * @start: start PFN to allocate
6711 * @end: one-past-the-last PFN to allocate
6712 * @migratetype: migratetype of the underlaying pageblocks (either
6713 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6714 * in range must have the same migratetype and it must
6715 * be either of the two.
6717 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6718 * aligned, however it's the caller's responsibility to guarantee that
6719 * we are the only thread that changes migrate type of pageblocks the
6722 * The PFN range must belong to a single zone.
6724 * Returns zero on success or negative error code. On success all
6725 * pages which PFN is in [start, end) are allocated for the caller and
6726 * need to be freed with free_contig_range().
6728 int alloc_contig_range(unsigned long start, unsigned long end,
6729 unsigned migratetype)
6731 unsigned long outer_start, outer_end;
6735 struct compact_control cc = {
6736 .nr_migratepages = 0,
6738 .zone = page_zone(pfn_to_page(start)),
6739 .mode = MIGRATE_SYNC,
6740 .ignore_skip_hint = true,
6742 INIT_LIST_HEAD(&cc.migratepages);
6745 * What we do here is we mark all pageblocks in range as
6746 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6747 * have different sizes, and due to the way page allocator
6748 * work, we align the range to biggest of the two pages so
6749 * that page allocator won't try to merge buddies from
6750 * different pageblocks and change MIGRATE_ISOLATE to some
6751 * other migration type.
6753 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6754 * migrate the pages from an unaligned range (ie. pages that
6755 * we are interested in). This will put all the pages in
6756 * range back to page allocator as MIGRATE_ISOLATE.
6758 * When this is done, we take the pages in range from page
6759 * allocator removing them from the buddy system. This way
6760 * page allocator will never consider using them.
6762 * This lets us mark the pageblocks back as
6763 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6764 * aligned range but not in the unaligned, original range are
6765 * put back to page allocator so that buddy can use them.
6768 ret = start_isolate_page_range(pfn_max_align_down(start),
6769 pfn_max_align_up(end), migratetype,
6774 ret = __alloc_contig_migrate_range(&cc, start, end);
6779 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6780 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6781 * more, all pages in [start, end) are free in page allocator.
6782 * What we are going to do is to allocate all pages from
6783 * [start, end) (that is remove them from page allocator).
6785 * The only problem is that pages at the beginning and at the
6786 * end of interesting range may be not aligned with pages that
6787 * page allocator holds, ie. they can be part of higher order
6788 * pages. Because of this, we reserve the bigger range and
6789 * once this is done free the pages we are not interested in.
6791 * We don't have to hold zone->lock here because the pages are
6792 * isolated thus they won't get removed from buddy.
6795 lru_add_drain_all();
6796 drain_all_pages(cc.zone);
6799 outer_start = start;
6800 while (!PageBuddy(pfn_to_page(outer_start))) {
6801 if (++order >= MAX_ORDER) {
6805 outer_start &= ~0UL << order;
6808 /* Make sure the range is really isolated. */
6809 if (test_pages_isolated(outer_start, end, false)) {
6810 pr_info("%s: [%lx, %lx) PFNs busy\n",
6811 __func__, outer_start, end);
6816 /* Grab isolated pages from freelists. */
6817 outer_end = isolate_freepages_range(&cc, outer_start, end);
6823 /* Free head and tail (if any) */
6824 if (start != outer_start)
6825 free_contig_range(outer_start, start - outer_start);
6826 if (end != outer_end)
6827 free_contig_range(end, outer_end - end);
6830 undo_isolate_page_range(pfn_max_align_down(start),
6831 pfn_max_align_up(end), migratetype);
6835 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6837 unsigned int count = 0;
6839 for (; nr_pages--; pfn++) {
6840 struct page *page = pfn_to_page(pfn);
6842 count += page_count(page) != 1;
6845 WARN(count != 0, "%d pages are still in use!\n", count);
6849 #ifdef CONFIG_MEMORY_HOTPLUG
6851 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6852 * page high values need to be recalulated.
6854 void __meminit zone_pcp_update(struct zone *zone)
6857 mutex_lock(&pcp_batch_high_lock);
6858 for_each_possible_cpu(cpu)
6859 pageset_set_high_and_batch(zone,
6860 per_cpu_ptr(zone->pageset, cpu));
6861 mutex_unlock(&pcp_batch_high_lock);
6865 void zone_pcp_reset(struct zone *zone)
6867 unsigned long flags;
6869 struct per_cpu_pageset *pset;
6871 /* avoid races with drain_pages() */
6872 local_irq_save(flags);
6873 if (zone->pageset != &boot_pageset) {
6874 for_each_online_cpu(cpu) {
6875 pset = per_cpu_ptr(zone->pageset, cpu);
6876 drain_zonestat(zone, pset);
6878 free_percpu(zone->pageset);
6879 zone->pageset = &boot_pageset;
6881 local_irq_restore(flags);
6884 #ifdef CONFIG_MEMORY_HOTREMOVE
6886 * All pages in the range must be isolated before calling this.
6889 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6893 unsigned int order, i;
6895 unsigned long flags;
6896 /* find the first valid pfn */
6897 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6902 zone = page_zone(pfn_to_page(pfn));
6903 spin_lock_irqsave(&zone->lock, flags);
6905 while (pfn < end_pfn) {
6906 if (!pfn_valid(pfn)) {
6910 page = pfn_to_page(pfn);
6912 * The HWPoisoned page may be not in buddy system, and
6913 * page_count() is not 0.
6915 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6917 SetPageReserved(page);
6921 BUG_ON(page_count(page));
6922 BUG_ON(!PageBuddy(page));
6923 order = page_order(page);
6924 #ifdef CONFIG_DEBUG_VM
6925 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6926 pfn, 1 << order, end_pfn);
6928 list_del(&page->lru);
6929 rmv_page_order(page);
6930 zone->free_area[order].nr_free--;
6931 for (i = 0; i < (1 << order); i++)
6932 SetPageReserved((page+i));
6933 pfn += (1 << order);
6935 spin_unlock_irqrestore(&zone->lock, flags);
6939 #ifdef CONFIG_MEMORY_FAILURE
6940 bool is_free_buddy_page(struct page *page)
6942 struct zone *zone = page_zone(page);
6943 unsigned long pfn = page_to_pfn(page);
6944 unsigned long flags;
6947 spin_lock_irqsave(&zone->lock, flags);
6948 for (order = 0; order < MAX_ORDER; order++) {
6949 struct page *page_head = page - (pfn & ((1 << order) - 1));
6951 if (PageBuddy(page_head) && page_order(page_head) >= order)
6954 spin_unlock_irqrestore(&zone->lock, flags);
6956 return order < MAX_ORDER;